Daily Papers

Modern AI benchmarks operate at a complexity that outpaces traditional verification methods. Tasks authored by domain experts often contain implicit assumptions, incomplete environment specifications, and brittle evaluation logic that human annotation cannot reliably catch. We introduce Auto Benchmark Audit (ABA), an agentic framework that systematically audits individual benchmark tasks, uncovering issues such as hidden environment dependencies, specification gaps, and limited grading logic. We run ABA on a collection of frontier LLM benchmarks and previous NeurIPS publications, totaling 168 benchmarks across nine domains. Across this corpus, ABA identifies critical issues including ambiguous task design, execution environment conflicts, and incorrect ground truths in over 25.7% of the evaluated tasks. The precision of these automated audits is validated by expert review and independent third-party reports such as upstream PRs. Crucially, we demonstrate that these problematic tasks severely distorts capability assessments for agents and LLMs: filtering out these tasks with issues shifts model rankings and increases average performance on SWE-bench Verified and Terminal-Bench 2 by 9.9% and 9.6%, respectively. We release the agentic tool and all task annotations to support the future development of frontier benchmarks.

Digital agents must often adapt to new environments whose interfaces, tools, and failure modes are not fully captured by pretrained models. We introduce RSIAgent, a training-free multi-agent framework for recursive self-improvement through autonomous memory construction. RSIAgent coordinates curriculum, actor, and verifier agents to continually explore the environment, validate outcomes, and retain environment-specific knowledge, including reusable causal relationships between actions, conditions, and consequences. It further adopts a broad-then-deep exploration strategy, combining parallel broad recursive self-exploration for discovering diverse environment structures with focused deep self-exploration for uncovering hard cases, hidden constraints, boundary conditions, and previously unknown causal dependencies. The resulting memory is frozen and can be directly reused for downstream tasks without updating model parameters. Experiments on OSWorld-v2 and Agent's Last Exam show that RSIAgent substantially improves strong open-source models, enabling Kimi-K3 and GLM-5.3 to outperform frontier closed-source models including GPT-6.

For the deployment of neural networks in resource-constrained environments, prior works have built lightweight architectures with convolution and attention for capturing local and global dependencies, respectively. Recently, the state space model has emerged as an effective global token interaction with its favorable linear computational cost in the number of tokens. Yet, efficient vision backbones built with SSM have been explored less. In this paper, we introduce Efficient Vision Mamba (EfficientViM), a novel architecture built on hidden state mixer-based state space duality (HSM-SSD) that efficiently captures global dependencies with further reduced computational cost. In the HSM-SSD layer, we redesign the previous SSD layer to enable the channel mixing operation within hidden states. Additionally, we propose multi-stage hidden state fusion to further reinforce the representation power of hidden states, and provide the design alleviating the bottleneck caused by the memory-bound operations. As a result, the EfficientViM family achieves a new state-of-the-art speed-accuracy trade-off on ImageNet-1k, offering up to a 0.7% performance improvement over the second-best model SHViT with faster speed. Further, we observe significant improvements in throughput and accuracy compared to prior works, when scaling images or employing distillation training. Code is available at https://github.com/mlvlab/EfficientViM.

Reinforcement Learning with Verifiable Rewards (RLVR) has recently emerged as a promising framework for improving reasoning abilities in Large Language Models (LLMs). However, policy optimized with binary verification prone to overlook potential valuable exploration in reasoning trajectory. In view of heavy annotation cost of golden Process Reward Models (PRMs), recent works attempt using auxiliary signals for reward shaping of process tokens, involving entropy and likelihood collected from logit space. In this work, we offer a novel perspective on shaping RLVR with flow rewards derived from latent space, and propose RLFR, where the flow fields of model latents are constructed from either off-policy high-quality data and on-policy rejection sampling data, and the velocity deviations of policy latents within it are quantified to serve as a reward signal. RLFR first demonstrates that a well-established flow field can be a sound environment for reward signal collection, highlighting the expressive latent space is much underexplored. Moreover, RLFR is able to compress any off-policy expert data as reference for constituting reward signals, and we show that the efficient context dependence compressed within the hidden states are utilized, rather than individual token-level denotation for context comprehending. Experiments on both language and multimodal reasoning benchmarks demonstrate the reliability of flow rewards, and suggesting a promising paradigm for reward shaping with auxiliary signals.

Modern LLM training pipelines increasingly rely on other models to generate data, filter corpora, judge outputs, and guide development decisions. These dependencies are recursive: a model may depend on an upstream artifact whose own dependencies are documented only in separate releases and artifacts. As a result, the full dependency structure is fragmented across heterogeneous public artifacts, with complexity and recursive depth far outpacing humans' ability to trace. We introduce ModSleuth, an agentic system that recursively reconstructs LLM dependency graphs from public artifacts with source-grounded evidence. We find that the primary challenge is no longer information extraction, but defining what constitutes a dependency and reconciling artifact references across inconsistent documentation. We address these challenges through a formalization that distinguishes direct and indirect dependencies, represents heterogeneous pipeline roles through operation-centered relationships, and resolves artifact identities across names, versions, and repositories. Applying ModSleuth to four public-artifact-rich LLM releases, we recover 1,060 source-verified dependencies and construct large-scale dependency graphs of modern LLM development. These graphs reveal multi-hop license obligations, train-evaluation coupling, discrepancies between released and training-time artifacts, and documentation inconsistencies that would otherwise be difficult to uncover. We release ModSleuth and the resulting dependency graphs to support transparent analysis of the increasingly complex ecosystems underlying modern LLMs.

Package dependencies are a critical control point in modern software supply chains. Dependency changes can substantially alter a project's security posture. As AI coding agents increasingly modify software via pull requests, it is unclear whether their dependency decisions introduce distinct security risks. We study 117,062 dependency changes from agent- and human-authored pull requests across seven ecosystems. Agents select known-vulnerable versions more often than humans (2.46% vs. 1.64%), and their vulnerable selections are more disruptive to remediate, with 36.8% requiring major-version upgrades compared to 12.9% for humans, despite patched alternatives existing in most cases. At the aggregate level, agent-driven dependency work yields a net vulnerability increase of 98, whereas human-authored work yields a net reduction of 1,316. These findings motivate pull-request-time vulnerability screening and registry-aware guardrails to make agent-driven dependency updates safer.

Pre-trained models (PTMs) are machine learning models that have been trained in advance, often on large-scale data, and can be reused for new tasks, thereby reducing the need for costly training from scratch. Their widespread adoption introduces a new class of software dependency, which we term Software Dependencies 2.0, extending beyond conventional libraries to learned behaviors embodied in trained models and their associated artifacts. The integration of PTMs as software dependencies in real projects remains unclear, potentially threatening maintainability and reliability of modern software systems that increasingly rely on them. Objective: In this study, we investigate Software Dependencies 2.0 in open-source software (OSS) projects by examining the reuse of PTMs, with a focus on how developers manage and integrate these models. Specifically, we seek to understand: (1) how OSS projects structure and document their PTM dependencies; (2) what stages and organizational patterns emerge in the reuse pipelines of PTMs within these projects; and (3) the interactions among PTMs and other learned components across pipeline stages. We conduct a mixed-methods analysis of a statistically significant random sample of 401 GitHub repositories from the PeaTMOSS dataset (28,575 repositories reusing PTMs from Hugging Face and PyTorch Hub). We quantitatively examine PTM reuse by identifying patterns and qualitatively investigate how developers integrate and manage these models in practice.

As LLMs move from text completion toward autonomous agents, they remain constrained by the standard chat interface, which lacks private working memory. This raises a fundamental question: can agents reliably perform interactive tasks that depend on hidden state? We define Private State Interactive Tasks (PSITs), which require agents to generate and maintain hidden information while producing public responses consistent with a fixed hidden state. We show theoretically that any agent restricted to the public conversation history cannot both keep the secret unresolved in the transcript and respond consistently with a fixed hidden state in PSITs, yielding an architectural impossibility theorem. To empirically validate this limitation, we introduce a self-consistency testing protocol that evaluates whether agents can maintain a hidden secret across forked dialogue branches. Standard chat-based LLMs and retrieval-based memory baselines fail this test regardless of scale, demonstrating that semantic retrieval does not enable true state maintenance. To address this, we propose a novel architecture incorporating an explicit private working memory; we demonstrate that this mechanism restores consistency with a fixed hidden state, establishing private state as a necessary component for PSIT-capable language agents. Our code is available at https://github.com/chandar-lab/Hangman

As Large Language Models (LLMs) are increasingly integrated into software development workflows, their trustworthiness has become a critical concern. However, in dependency recommendation scenarios, the reliability of LLMs is undermined by widespread package hallucinations, where models often recommend hallucinated packages. Recent studies have proposed a range of approaches to mitigate this issue. Nevertheless, existing approaches typically merely reduce hallucination rates rather than eliminate them, leaving persistent software security risks. In this work, we argue that package hallucinations are theoretically preventable based on the key insight that package validity is decidable through finite and enumerable authoritative package lists. Building on this, we propose PackMonitor, the first approach capable of fundamentally eliminating package hallucinations by continuously monitoring the model's decoding process and intervening when necessary. To implement this in practice, PackMonitor addresses three key challenges: (1) determining when to trigger intervention via a Context-Aware Parser that continuously monitors model outputs and selectively activates intervening only during installation command generation; (2) resolving how to intervene by employing a Package-Name Intervenor that strictly limits the decoding space to an authoritative package list; and (3) ensuring monitoring efficiency through a DFA-Caching Mechanism that enables scalability to millions of packages with negligible overhead. Extensive experiments on five widely used LLMs demonstrate that PackMonitor is a training-free, plug-and-play solution that consistently reduces package hallucination rates to zero while maintaining low-latency inference and preserving original model capabilities.

As LLM-based agents increasingly rely on external tools, it is important to evaluate their ability to sustain tool-grounded reasoning beyond familiar workflows and short-range interactions. We introduce AgentEscapeBench, an escape-room-style benchmark that tests whether agents can infer, execute, and revise novel tool-use procedures under explicit long-range dependency constraints. Each task defines a directed acyclic dependency graph over tools and items, requiring agents to invoke real external functions, track hidden state revealed incrementally, propagate intermediate results, and submit a deterministically verifiable final answer. AgentEscapeBench includes 270 instances across five difficulty tiers and supports fully automated evaluation. Experiments with sixteen LLM agents and human participants show that performance drops sharply as dependency depth increases: humans decline from 98.3% success at difficulty-5 to 80.0% at difficulty-25, while the best model drops from 90.0% to 60.0%. Trajectory analysis attributes model failures mainly to breakdowns in long-range state tracking, clue adherence, and intermediate-result propagation. These findings suggest that current agents can often handle local tool use but still struggle with deep contextual dependencies. We hope AgentEscapeBench can serve as a diagnostic testbed for measuring current agent capabilities and informing future training efforts toward more robust general-purpose reasoning, action, and adaptation.

Hidden chain-of-thought (CoT) traces, especially those from frontier proprietary large reasoning models (LRMs), are valuable model assets. Yet whether these hidden CoTs can be directly extracted from black-box models remains largely unexplored. In this work, we systematically study whether hidden CoTs can be extracted near-verbatim from black-box LRMs through API interactions. We identify a previously overlooked reasoning replay surface between tool calls and develop EchoCoT, a multi-step attack that iteratively extracts hidden CoTs using API-returned fidelity signals. We further develop an LLM-based optimization framework that automatically searches for an effective universal injection trajectory across various datasets. We evaluate EchoCoT on three open-source and five frontier proprietary LRMs. On open-source LRMs, EchoCoT achieves up to 66.4\% near-verbatim extraction success, with the extracted trace length within 10\% of the target and at least 90\% of tokens exactly matching the target CoT. The same injection trajectory also generalizes to unseen datasets, achieving up to 80\% extraction success under the same criterion. For tested frontier proprietary LRMs, a substantial fraction of extracted CoTs closely align with provider-reported reasoning lengths and available CoT summaries. EchoCoT can also extract very long CoTs: on Gemini-2.5, it extracts 33,463 tokens from a 32,948-token target. These results establish hidden-CoT extraction as a practical security risk and highlight the need to better protect hidden CoT assets.

Retrieving task-relevant code dependencies is critical for repository-level code generation. Existing approaches obtain dependency context through similarity-based retrieval, repository graphs, or LLM-driven exploration. A central challenge is to identify the dispersed dependencies needed to implement a functional requirement, rather than merely retrieve broadly related code. This paper presents FeatLens, a feature-guided dependency retrieval approach for repository-level code generation. FeatLens first builds an offline index linking natural-language feature descriptions to function-level code entities. Given a task requirement and target function location, it selects relevant features and constructs dependency seeds. It then expands the seed graph and applies semantic and structural heuristics to select dependencies for inclusion in the generation context. Experiments on DevEval and EvoCodeBench show that FeatLens achieves the highest dependency recall among the evaluated baselines at each fixed retrieval budget, reaching DR@15 values of 0.527 and 0.529, respectively. On DevEval, the retrieved context supports the highest non-oracle Pass@1 with both DeepSeek-V3.2 and GPT-5-mini. These results are obtained without LLM calls in the online dependency retrieval stage.

Open-Source Software (OSS) vulnerabilities bring great challenges to the software security and pose potential risks to our society. Enormous efforts have been devoted into automated vulnerability detection, among which deep learning (DL)-based approaches have proven to be the most effective. However, the current labeled data present the following limitations: (1) Tangled Patches: Developers may submit code changes unrelated to vulnerability fixes within patches, leading to tangled patches. (2) Lacking Inter-procedural Vulnerabilities: The existing vulnerability datasets typically contain function-level and file-level vulnerabilities, ignoring the relations between functions, thus rendering the approaches unable to detect the inter-procedural vulnerabilities. (3) Outdated Patches: The existing datasets usually contain outdated patches, which may bias the model during training. To address the above limitations, in this paper, we propose an automated data collection framework and construct the first repository-level high-quality vulnerability dataset named ReposVul. The proposed framework mainly contains three modules: (1) A vulnerability untangling module, aiming at distinguishing vulnerability-fixing related code changes from tangled patches, in which the Large Language Models (LLMs) and static analysis tools are jointly employed. (2) A multi-granularity dependency extraction module, aiming at capturing the inter-procedural call relationships of vulnerabilities, in which we construct multiple-granularity information for each vulnerability patch, including repository-level, file-level, function-level, and line-level. (3) A trace-based filtering module, aiming at filtering the outdated patches, which leverages the file path trace-based filter and commit time trace-based filter to construct an up-to-date dataset.

AI agents are increasingly created inside organizations by non-engineering users through low-code, no-code, and conversational development environments. This democratization enables rapid local innovation, but it also creates a reliability gap: agents that appear to users as simple productivity artifacts may depend on changing models, tools, retrieval sources, permissions, prompts, schedules, and external services. These dependencies can cause silent degradation long after deployment, even when no user directly modifies the agent. This paper identifies the reliability challenge created by democratized AI agent creation and proposes a lightweight continuous-assurance framework for citizen-created organizational agents. The framework combines dependency mapping, readiness contracts, scheduled checks, diagnostics, and lifecycle governance to assess whether an agent remains operationally ready under expected conditions. We also present an initial prototype auditor and scenario-based assessment showing how the proposed taxonomy can be translated into practical checks and actionable remediation guidance.

You have an environment, a model, and a reinforcement learning library that are designed to work together but don't. PufferLib makes them play nice. The library provides one-line environment wrappers that eliminate common compatibility problems and fast vectorization to accelerate training. With PufferLib, you can use familiar libraries like CleanRL and SB3 to scale from classic benchmarks like Atari and Procgen to complex simulators like NetHack and Neural MMO. We release pip packages and prebuilt images with dependencies for dozens of environments. All of our code is free and open-source software under the MIT license, complete with baselines, documentation, and support at pufferai.github.io.

Software ecosystems rely on centralized package registries, such as Maven, to enable code reuse and collaboration. However, the interconnected nature of these ecosystems amplifies the risks posed by security vulnerabilities in direct and transitive dependencies. While numerous studies have examined vulnerabilities in Maven and other ecosystems, there remains a gap in understanding the behavior of vulnerabilities across parent and dependent packages, and the response times of maintainers in addressing vulnerabilities. This study analyzes the lifecycle of 3,362 CVEs in Maven to uncover patterns in vulnerability mitigation and identify factors influencing at-risk packages. We conducted a comprehensive study integrating temporal analyses of CVE lifecycles, correlation analyses of GitHub repository metrics, and assessments of library maintainers' response times to patch vulnerabilities, utilizing a package dependency graph for Maven. A key finding reveals a trend in "Publish-Before-Patch" scenarios: maintainers prioritize patching severe vulnerabilities more quickly after public disclosure, reducing response time by 48.3% from low (151 days) to critical severity (78 days). Additionally, project characteristics, such as contributor absence factor and issue activity, strongly correlate with the presence of CVEs. Leveraging tools such as the Goblin Ecosystem, OSV.dev, and OpenDigger, our findings provide insights into the practices and challenges of managing security risks in Maven.

LLM agents learn by interacting with environments, yet these environments are hand-built and static: blind to an agent's weaknesses, and quickly left behind as it improves. While recent environment generation methods attempt to address this, they require domain-specific pipelines, rely on expensive or unreliable verifiers, and still produce static environments. To alleviate the engineering burden of rebuilding environments from scratch, we propose Environment Harness (EnvHarness), a programmable layer of plug-in components that wraps a static environment to reshape its behavior without modifying the underlying logic. Operating through standard interfaces, EnvHarness applies across diverse domains while ensuring every reshaped environment retains its original verifier. To automate this process, we introduce EnvRigger, which treats the target policy as a black box, observing its execution trajectories to synthesize EnvHarness components targeting diagnosed flaws, and validating them via fresh rollouts. Across five benchmarks in four domains, EnvHarness outperforms both original environments and domain-specific environment generation pipelines, achieving up to a 9.0-point improvement on held-out instances with 9.8% fewer execution steps. Furthermore, EnvHarness provides a superior optimization signal for reinforcement learning, enabling continuous, targeted co-evolution of the policy and its environment.

LLM agents excel when environments are mostly static and the needed information fits in a model's context window, but they often fail in open-ended investigations where explanations must be constructed by iteratively mining evidence from massive, heterogeneous operational data. These investigations exhibit hidden dependency structure: entities interact, signals co-vary, and the importance of a fact may only become clear after other evidence is discovered. Because the context window is bounded, agents must summarize intermediate findings before their significance is known, increasing the risk of discarding key evidence. ReAct-style agents are especially brittle in this regime. Their retrieve-summarize-reason loop makes conclusions sensitive to exploration order and introduces run-to-run non-determinism, producing a reliability gap where Pass-at-k may be high but Majority-at-k remains low. Simply sampling more rollouts or generating longer reasoning traces does not reliably stabilize results, since hypotheses cannot be autonomously checked as new evidence arrives and there is no explicit mechanism for belief bookkeeping and revision. In addition, ReAct entangles semantic reasoning with controller duties such as tool orchestration and state tracking, so execution errors and plan drift degrade reasoning while consuming scarce context. We address these issues by formulating investigation as abductive reasoning over a dependency graph and proposing EoG (Explanations over Graphs), a disaggregated framework in which an LLM performs bounded local evidence mining and labeling (cause vs symptom) while a deterministic controller manages traversal, state, and belief propagation to compute a minimal explanatory frontier. On a representative ITBench diagnostics task, EoG improves both accuracy and run-to-run consistency over ReAct baselines, including a 7x average gain in Majority-at-k entity F1.

AI coding agents set up projects by reading documentation and installing the dependencies it lists, without verifying their names, sources, or known vulnerabilities. By editing only a README, requirements file, or Makefile, an attacker can redirect the agent to an untrusted registry, a known-vulnerable version, or a wrong-but-plausible name: documentation becomes a vector for code execution. We present the first systematic evaluation of package-install-time supply-chain attacks delivered through ordinary project-setup documentation across production coding-agent harnesses, probing frontier models on twelve scenarios in five attack classes, grounded in documented incidents. The same model catches an attack through one harness and installs it through another: install-time security rests on the harness-model combination, not the model alone. Agents catch blatant typosquats reliably, but plausible separator-confusion names (azurecore for azure-core) slip through, and how often depends on the harness-model pairing. Source-based attacks like registry redirection are missed almost everywhere. The source blind spot recurs on npm and Cargo, where nearly every model installs the untrusted dependency; name detection carries over less consistently across ecosystems. Security-oriented prompts recover part of the gap but only for the dimension they name; a deterministic pre-install check that verifies names, sources, and versions before any code runs closes most of it.

While large language models (LLMs) have shown considerable promise in code generation, real-world software development demands advanced repository-level reasoning. This includes understanding dependencies, project structures, and managing multi-file changes. However, the ability of LLMs to effectively comprehend and handle complex code repositories has yet to be fully explored. To address challenges, we introduce a hierarchical benchmark designed to evaluate repository dependency understanding (DependEval). Benchmark is based on 15,576 repositories collected from real-world websites. It evaluates models on three core tasks: Dependency Recognition, Repository Construction, and Multi-file Editing, across 8 programming languages from actual code repositories. Our evaluation of over 25 LLMs reveals substantial performance gaps and provides valuable insights into repository-level code understanding.

The automatic detection of software vulnerabilities is an important research problem. However, existing solutions to this problem rely on human experts to define features and often miss many vulnerabilities (i.e., incurring high false negative rate). In this paper, we initiate the study of using deep learning-based vulnerability detection to relieve human experts from the tedious and subjective task of manually defining features. Since deep learning is motivated to deal with problems that are very different from the problem of vulnerability detection, we need some guiding principles for applying deep learning to vulnerability detection. In particular, we need to find representations of software programs that are suitable for deep learning. For this purpose, we propose using code gadgets to represent programs and then transform them into vectors, where a code gadget is a number of (not necessarily consecutive) lines of code that are semantically related to each other. This leads to the design and implementation of a deep learning-based vulnerability detection system, called Vulnerability Deep Pecker (VulDeePecker). In order to evaluate VulDeePecker, we present the first vulnerability dataset for deep learning approaches. Experimental results show that VulDeePecker can achieve much fewer false negatives (with reasonable false positives) than other approaches. We further apply VulDeePecker to 3 software products (namely Xen, Seamonkey, and Libav) and detect 4 vulnerabilities, which are not reported in the National Vulnerability Database but were "silently" patched by the vendors when releasing later versions of these products; in contrast, these vulnerabilities are almost entirely missed by the other vulnerability detection systems we experimented with.

As terminal-based code agents become prevalent, agent trajectories have accumulated at scale, while realistic, executable environments remain scarce. However, environments are what agent post-training actually requires: each can be re-queried into many verifiable tasks and provides execution feedback, whereas a trajectory is a single frozen demonstration. Rather than generating environments from scratch, we observe that the tool-execution history in existing trajectories exposes the structure and contents of the environments in which they ran, making it possible to reconstruct those environments from the trajectories themselves. Thus, we introduce Terminal-Universe, a framework which turns each trajectory into a reusable environment and explores it for synthesizing new tasks and continued interactions. Specifically, Terminal-Universe replays the file operations recorded in a trajectory to restore each file before the agent modified it, yielding a partial workspace; a completion agent then supplies the missing files and dependencies. On this recovered workspace, we both reconstruct the original intent task and synthesize entirely new ones. Besides, we also scale the tasks along two complementary axes: breadth and depth. For breadth, we mine directional dependency relations between related environments and synthesize cross-workspace queries spanning multiple codebases, as developers routinely do in real-world development. For depth, we extend the initial single-turn query into a multi-round session that captures iterative user feedback and requirement refinement via a user agent. Applied to public terminal agent trajectories, Terminal-Universe produces 37.3k task-sufficient environments. Supervised fine-tuning of Qwen3.5-27B on this corpus improves single-round performance on Terminal-Bench 2.1 by 11.9 points and multi-round performance on EvoCode-Bench v2 MT@4 by 13.8 points.

As LLMs are deployed as agents, reliable monitoring requires knowing not only what they output, but which instructions are steering their behavior. This is difficult when models infer unintended subgoals, follow contextual cues, or are influenced by prompt injections and hidden objectives. While activation-to-language methods suggest that hidden states can reveal natural-language information, existing approaches are not designed to recover the full set of simultaneous instructions, constraints, prohibitions, and subgoals active in agentic settings. We formalize this problem as instruction set retrieval and introduce PRISM, an activation-conditioned interpreter that decodes hidden states from a frozen target model into a faithful bullet list of active instructions. Unlike prior activation-to-language methods, PRISM is trained to recover instruction sets directly, using judge-guided GRPO to reward covered instructions and penalize unsupported ones. Across benign, constrained, prompt-injection, and hidden-objective settings, PRISM outperforms activation-to-language baselines, especially on security-relevant objectives.

Large language models (LLMs) are increasingly deployed as purpose-specific agents to handle domain-specific tasks such as customer service and code generation. These agents are expected to comply with not only generic safety guardrails but also purpose-specific restrictions tailored to their designated roles. Such additional restrictions enlarge the attack surface, particularly to prompt injection (PI) attacks. To defend against such attacks, existing detection methods primarily rely on analyzing input-output patterns, yet yield limited effectiveness. To address this limitation, we turn to analyzing the hidden activation space and discover that LLMs inherently retain latent policy-violation (PV) concepts when prompted with requests beyond their designated purpose. Particularly, PV concepts capture the semantics of conflicts between user queries and predefined restrictions, implicitly reflecting LLMs' intrinsic awareness of recognizing policy violations. Building on this insight, we propose PVDetector, a training-free framework that detects PI attacks during LLM inference by measuring hidden-state alignment with PV concepts, which are derived offline from the contrastive pairs of policy-violating and policy-compliant prompts. Experiments across multiple LLMs and datasets show that PVDetector achieves <1\% false negative rate with minimal auxiliary overhead, consistently outperforming state-of-the-art methods. Our code is available at https://github.com/Claresigle/PVDetector .

Hallucinations, outputs that sound plausible but are factually incorrect, remain an open challenge for deployed LLMs. In code generation, models frequently hallucinate non-existent software packages, recommending imports and installation commands for fictional libraries. This creates a critical supply-chain vulnerability: an attacker can proactively register such packages on public registries with malicious payloads that are subsequently installed and executed by developers or autonomous agents, a class of package confusion attack known as slopsquatting. Once a model is deployed, mitigating this failure mode is difficult: full retraining is costly, and existing approaches either cause severe degradation of model utility or rely on a pre-specified forget-set, an assumption that does not apply to the unbounded space of hallucinations. To address this problem, we present Adaptive Unlearning (AU), a post-deployment framework that surgically suppresses hallucinations while preserving general model utility. AU introduces a hybrid token-level objective that simultaneously reinforces valid outputs and suppresses hallucinated ones. Combined with an adaptive discovery loop that continuously surfaces new hallucination-inducing contexts without human supervision, AU enables generalization to unseen prompts and hallucinations. We demonstrate that AU reduces package hallucination rates by 81%, corresponding to a substantial reduction in slopsquatting attack surface, while maintaining performance on standard coding benchmarks. Our analysis shows that distributional changes are concentrated on package-related generations, leaving general coding behavior largely unaffected and confirming that AU's effect is isolated to the targeted distribution. AU operates entirely on model-generated data, requires no human annotation, and generalizes across domains.

Vision-Language Models (VLMs) are increasingly deployed as autonomous agents to navigate mobile graphical user interfaces (GUIs). Operating in dynamic on-device ecosystems, which include notifications, pop-ups, and inter-app interactions, exposes them to a unique and underexplored threat vector: environmental injection. Unlike prompt-based attacks that manipulate textual instructions, environmental injection corrupts an agent's visual perception by inserting adversarial UI elements (for example, deceptive overlays or spoofed notifications) directly into the GUI. This bypasses textual safeguards and can derail execution, causing privacy leakage, financial loss, or irreversible device compromise. To systematically evaluate this threat, we introduce GhostEI-Bench, the first benchmark for assessing mobile agents under environmental injection attacks within dynamic, executable environments. Moving beyond static image-based assessments, GhostEI-Bench injects adversarial events into realistic application workflows inside fully operational Android emulators and evaluates performance across critical risk scenarios. We further propose a judge-LLM protocol that conducts fine-grained failure analysis by reviewing the agent's action trajectory alongside the corresponding screenshot sequence, pinpointing failure in perception, recognition, or reasoning. Comprehensive experiments on state-of-the-art agents reveal pronounced vulnerability to deceptive environmental cues: current models systematically fail to perceive and reason about manipulated UIs. GhostEI-Bench provides a framework for quantifying and mitigating this emerging threat, paving the way toward more robust and secure embodied agents.

Recent advances in Large Language Models (LLMs) have enabled researchers to focus on practical repository-level tasks in software engineering domain. In this work, we consider a cornerstone task for automating work with software repositories-environment setup, i.e., a task of configuring a repository-specific development environment on a system. Existing studies on environment setup introduce innovative agentic strategies, but their evaluation is often based on small datasets that may not capture the full range of configuration challenges encountered in practice. To address this gap, we introduce a comprehensive environment setup benchmark EnvBench. It encompasses 329 Python and 665 JVM-based (Java, Kotlin) repositories, with a focus on repositories that present genuine configuration challenges, excluding projects that can be fully configured by simple deterministic scripts. To enable further benchmark extension and usage for model tuning, we implement two automatic metrics: a static analysis check for missing imports in Python and a compilation check for JVM languages. We demonstrate the applicability of our benchmark by evaluating three environment setup approaches, including a simple zero-shot baseline and two agentic workflows, that we test with two powerful LLM backbones, GPT-4o and GPT-4o-mini. The best approach manages to successfully configure 6.69% repositories for Python and 29.47% repositories for JVM, suggesting that EnvBench remains challenging for current approaches. Our benchmark suite is publicly available at https://github.com/JetBrains-Research/EnvBench. The dataset and experiment trajectories are available at https://jb.gg/envbench.

We present the first empirical study of AI-generated pull requests that are 'silent,' meaning no comments or discussions accompany them. This absence of any comments or discussions associated with such silent AI pull requests (SPRs) poses a unique challenge in understanding the rationale for their acceptance or rejection. Hence, we quantitatively study 4,762 SPRs of five AI agents made to popular Python repositories drawn from the AIDev public dataset. We examine SPRs impact on code complexity, other quality issues, and security vulnerabilities, especially to determine whether these insights can hint at the rationale for acceptance or rejection of SPRs.

We present RepoST, a scalable method to construct environments that provide execution feedback for repository-level code generation for both training and evaluation. Unlike existing works that aim to build entire repositories for execution, which is challenging for both human and LLMs, we provide execution feedback with sandbox testing, which isolates a given target function and its dependencies to a separate script for testing. Sandbox testing reduces the complexity of external dependencies and enables constructing environments at a large scale. We use our method to construct RepoST-Train, a large-scale train set with 7,415 functions from 832 repositories. Training with the execution feedback provided by RepoST-Train leads to a performance gain of 5.5% Pass@1 on HumanEval and 3.5% Pass@1 on RepoEval. We also build an evaluation dataset, RepoST-Eval, and benchmark 12 code generation models.

Automatically compiling open-source software (OSS) projects is a vital, labor-intensive, and complex task, which makes it a good challenge for LLM Agents. Existing methods rely on manually curated rules and workflows, which cannot adapt to OSS that requires customized configuration or environment setup. Recent attempts using Large Language Models (LLMs) used selective evaluation on a subset of highly rated OSS, a practice that underestimates the realistic challenges of OSS compilation. In practice, compilation instructions are often absent, dependencies are undocumented, and successful builds may even require patching source files or modifying build scripts. We propose a more challenging and realistic benchmark, BUILD-BENCH, comprising OSS that are more diverse in quality, scale, and characteristics. Furthermore, we propose a strong baseline LLM-based agent, OSS-BUILD-AGENT, an effective system with enhanced build instruction retrieval module that achieves state-of-the-art performance on BUILD-BENCH and is adaptable to heterogeneous OSS characteristics. We also provide detailed analysis regarding different compilation method design choices and their influence to the whole task, offering insights to guide future advances. We believe performance on BUILD-BENCH can faithfully reflect an agent's ability to tackle compilation as a complex software engineering tasks, and, as such, our benchmark will spur innovation with a significant impact on downstream applications in the fields of software development and software security.

Trusted Execution Environments (TEEs) provide hardware-based isolation to protect sensitive data and computations from potentially compromised operating systems (OS). However, TEE applications inevitably interact with the untrusted OS through SDK interfaces, and improper partitioning can introduce severe vulnerabilities such as data leakage and code injection. While prior work has proposed static analysis tools to detect such issues, automated repair remains largely unexplored. This problem is particularly challenging due to three TEE-specific factors: the lack of standardized secure development guidelines, the difficulty of extracting semantic information from low-level C code, and the absence of mature testing and validation methods. In this work, we present TEERepair, a framework for automatically repairing bad partitioning issues in TEE applications. Our approach tackles the above challenges by introducing a domain-specific language (DSL) to encode repair rules that express and capture common TEE security patterns, which are instantiated as patch templates with placeholders for context-specific variables. We then leverage large language models (LLMs) to reason about code semantics and synthesize context-aware patches, and further generate test clients to validate the repairs. We evaluate TEERepair on the TEE Partitioning Errors Benchmark (PartitioningE-Bench), achieving a significantly higher repair success rate of 87.6% compared to baselines. Furthermore, applying TEERepair to real-world TEE projects, we submitted 5 repair pull requests, 2 of which have been confirmed and merged by project maintainers.

Terminal agents are increasingly capable of executing complex, long-horizon tasks autonomously from a single user prompt. To do so, they must interpret instructions encountered in the environment (e.g., README files, code comments, stack traces) and determine their relevance to the task. This creates a fundamental challenge: relevant cues must be followed to complete a task, whereas irrelevant or misleading ones must be ignored. Existing benchmarks do not capture this ability. An agent may appear capable by blindly following all instructions, or appear robust by ignoring them altogether. We introduce TAB (Task Alignment Benchmark), a suite of 89 terminal tasks derived from Terminal-Bench 2.1. Each task is intentionally underspecified, with missing information provided as a necessary cue embedded in a natural environmental artifact, alongside a plausible but irrelevant distractor. Solving these tasks requires selectively using the cue while ignoring the distractor. Applying TAB to ten frontier agents reveals a systematic gap between task capability and task alignment. The strongest Terminal-Bench agent achieves high task completion but low task alignment on TAB. Evaluating six prompt-injection defenses further shows that suppressing distractor execution also suppresses the cues required for task completion. These results demonstrate that task-aligned agents require selective use of environmental instructions rather than blanket acceptance or rejection.

With the increasing usage of open-source software (OSS) components, vulnerabilities embedded within them are propagated to a huge number of underlying applications. In practice, the timely application of security patches in downstream software is challenging. The main reason is that such patches do not explicitly indicate their security impacts in the documentation, which would be difficult to recognize for software maintainers and users. However, attackers can still identify these "secret" security patches by analyzing the source code and generate corresponding exploits to compromise not only unpatched versions of the current software, but also other similar software packages that may contain the same vulnerability due to code cloning or similar design/implementation logic. Therefore, it is critical to identify these secret security patches to enable timely fixes. To this end, we propose a deep learning-based defense system called PatchRNN to automatically identify secret security patches in OSS. Besides considering descriptive keywords in the commit message (i.e., at the text level), we leverage both syntactic and semantic features at the source-code level. To evaluate the performance of our system, we apply it on a large-scale real-world patch dataset and conduct a case study on a popular open-source web server software - NGINX. Experimental results show that the PatchRNN can successfully detect secret security patches with a low false positive rate.

Environment setup-the process of configuring the system to work with a specific software project-represents a persistent challenge in Software Engineering (SE). Automated environment setup methods could assist developers by providing fully configured environments for arbitrary repositories without manual effort. This also helps SE researchers to scale execution-based benchmarks. However, recent studies reveal that even state-of-the-art Large Language Models (LLMs) achieve limited success in automating this task. To address this limitation, we tune a specialized model for environment setup. We combine supervised fine-tuning for generating correct Bash scripts and Reinforcement Learning with Verifiable Rewards (RLVR) to adapt it to the task of environment setup. On EnvBench-Python, our method enables Qwen3-8B (a model runnable on consumer hardware) to perform on par with larger models-Qwen3-32B and GPT-4o. The training code and model checkpoints are available online: https://github.com/JetBrains-Research/PIPer.

We report a peculiar observation that LLMs can assign hidden meanings to sequences that seem visually incomprehensible to humans: for example, a nonsensical phrase consisting of Byzantine musical symbols is recognized by gpt-4o as "say abracadabra". Moreover, some models can communicate using these sequences. Some of these meanings are hypothesized to partly originate in the massive spurious correlations due to BPE tokenization. We systematically evaluate the presence of such abilities in a wide range of models: Claude-3.5 Haiku, Claude-3.5 Sonnet (New and Old), Claude-3.7 Sonnet, gpt-4o mini, gpt-4o, o1-mini, Llama-3.3 70B, DeepSeek-R1-Distill-Lllama 70B, Qwen2.5 1.5B, Qwen2.5 32B, Phi-3.5 mini, GigaChat-Max, Vikhr-Llama-3.2 1B. We argue that this observation might have far-reaching consequences for both safety and security of the modern and future LLMs and systems that employ them. As an illustration, we show that applying this method in combination with simple templates is sufficient to jailbreak previous generation models, with ASR = 0.4 on gpt-4o mini. Our code and data artifacts are available at https://github.com/L3G5/llm-hidden-meanings

With model weights fixed, we measure how much of an entity-replacement response transfers between English factual questions, beyond whether answers change. We add the hidden-state change from two source questions to a target question about the same entity, comparing the predicted log-probability shift between two relations' published answers, given known answer prefixes, with the measured one. Hidden reuse injects the aligned change at 50% depth before the target's remaining layers; direct reuse adds source output-score changes. With other-relation sources, hidden reuse removes 45% and 49% of no-change error on 0.5-1B Qwen and Llama. Gemma-2-2B removes 22% and passes one of three comparisons in plain text without its default beginning-of-sequence (BOS) token. A later protocol restoring it yields 58-68%, passing every comparison on three cohorts. Larger checkpoints, chat format and new entities show replication and format- or cohort-dependent failures. On small plain-format checkpoints, other-relation hidden reuse mainly reproduces the shift both relations' questions share, without reliably lowering relation-selective error. With target-relation sources, hidden reuse's post hoc advantage over direct reuse is 0.2-0.9 nats outside chat format, unconfirmed and not replacement-specific. The same relation-neutral change at both questions' entity tokens lowers relation-selective error more than random or other-replacement changes, indicating receiving-question dependence.

LLM agents often rely on Skills to describe available tools and recommended procedures. We study a hidden-comment prompt injection risk in this documentation layer: when a Markdown Skill is rendered to HTML, HTML comment blocks can become invisible to human reviewers, yet the raw text may still be supplied verbatim to the model. In experiments, we find that DeepSeek-V3.2 and GLM-4.5-Air can be influenced by malicious instructions embedded in a hidden comment appended to an otherwise legitimate Skill, yielding outputs that contain sensitive tool intentions. A short defensive system prompt that treats Skills as untrusted and forbids sensitive actions prevents these malicious tool calls and instead surfaces the suspicious hidden instructions.

Unlearning in large language models (LLMs) is typically evaluated at the output level, where a model appears to suppress sensitive or undesirable content. In this work, we show that such evaluations can create an illusion of forgetting: even when output-level leakage is eliminated, sensitive information can remain encoded in the model's hidden representations. We first provide a theoretical analysis establishing a fundamental separation between output suppression and representational erasure. Specifically, we show that the decoder can be made arbitrarily insensitive to sensitive directions, driving output-level leakage to zero, while the hidden representations retain the underlying information. To empirically validate this phenomenon, we train generative probe decoders on hidden states across transformer layers, enabling layer-wise measurement of information leakage. Across three widely used benchmarks, TOFU, MUSE, and WMDP, and state-of-the-art unlearning methods, we find that substantial sensitive information remains recoverable from hidden representations, even when standard output-level metrics indicate successful unlearning. To address this gap, we propose Probe-Adversarial Representation Suppression (PARS), an unlearning objective that adversarially minimizes the extractable information from hidden representations. PARS directly targets representational leakage and provides significantly stronger guarantees of erasure under adversarial probing and relearning attacks, outperforming all evaluated baselines. Our results highlight a fundamental limitation of existing unlearning paradigms and suggest that true forgetting in LLMs requires controlling not only model outputs, but also the information encoded in hidden representations. Codes are available at https://github.com/OptimAI-Lab/HiddenStateUnlearning.

Large language models (LLMs) increasingly act as autonomous agents, using tools to execute code, read and write files, and access networks, creating novel security risks. To mitigate these risks, agents are commonly deployed and evaluated in isolated "sandbox" environments, often implemented using Docker/OCI containers. We introduce SANDBOXESCAPEBENCH, an open benchmark that safely measures an LLM's capacity to break out of these sandboxes. The benchmark is implemented as an Inspect AI Capture the Flag (CTF) evaluation utilising a nested sandbox architecture with the outer layer containing the flag and no known vulnerabilities. Following a threat model of a motivated adversarial agent with shell access inside a container, SANDBOXESCAPEBENCH covers a spectrum of sandboxescape mechanisms spanning misconfiguration, privilege allocation mistakes, kernel flaws, and runtime/orchestration weaknesses. We find that, when vulnerabilities are added, LLMs are able to identify and exploit them, showing that use of evaluation like SANDBOXESCAPEBENCH is needed to ensure sandboxing continues to provide the encapsulation needed for highly-capable models.

Large language models can hide hidden behaviors that activate only under narrow conditions, such as backdoor triggers, sleeper-agent deployment cues, sandbagging, or topic-conditioned censorship. Such behaviors are difficult to detect without prior knowledge what to look for. We present activation-matched finetuning, an unsupervised detection method that assumes no knowledge of the trigger or the target behavior. Given a suspect model and a publicly available anchor, we finetune the anchor to reproduce the suspect's activations on a small benign corpus, and score each evaluation prompt by the residual between the two models. Since no benign corpus covers the sparse trigger region, the reference learns the benign computation but not the hidden behavior. Therefore, trigger prompts -- and, crucially, their semantic neighbors -- incur a large residual that signal the presence of unusual behavior to the defender. Testing our method across third-party models and custom models, activation-matched finetuning surfaces hidden behavior reliably. Furthermore, we empirically consider a natural defense-aware attack and showcase that it fails to suppress our detection method without sacrificing the behavior itself.

Many works have recently proposed the use of Large Language Model (LLM) based agents for performing `repository level' tasks, loosely defined as a set of tasks whose scopes are greater than a single file. This has led to speculation that the orchestration of these repository-level tasks could lead to software engineering agents capable of performing almost independently of human intervention. However, of the suite of tasks that would need to be performed by this autonomous software engineering agent, we argue that one important task is missing, which is to fulfil project level dependency by installing other repositories. To investigate the feasibility of this repository level installation task, we introduce a benchmark of of repository installation tasks curated from 40 open source Python projects, which includes a ground truth installation process for each target repository. Further, we propose Installamatic, an agent which aims to perform and verify the installation of a given repository by searching for relevant instructions from documentation in the repository. Empirical experiments reveal that that 55% of the studied repositories can be automatically installed by our agent at least one out of ten times. Through further analysis, we identify the common causes for our agent's inability to install a repository, discuss the challenges faced in the design and implementation of such an agent and consider the implications that such an agent could have for developers.

The integration of additional modalities increases the susceptibility of large vision-language models (LVLMs) to safety risks, such as jailbreak attacks, compared to their language-only counterparts. While existing research primarily focuses on post-hoc alignment techniques, the underlying safety mechanisms within LVLMs remain largely unexplored. In this work , we investigate whether LVLMs inherently encode safety-relevant signals within their internal activations during inference. Our findings reveal that LVLMs exhibit distinct activation patterns when processing unsafe prompts, which can be leveraged to detect and mitigate adversarial inputs without requiring extensive fine-tuning. Building on this insight, we introduce HiddenDetect, a novel tuning-free framework that harnesses internal model activations to enhance safety. Experimental results show that {HiddenDetect} surpasses state-of-the-art methods in detecting jailbreak attacks against LVLMs. By utilizing intrinsic safety-aware patterns, our method provides an efficient and scalable solution for strengthening LVLM robustness against multimodal threats. Our code will be released publicly at https://github.com/leigest519/HiddenDetect.

This work presents LURK (Latent UnleaRned Knowledge), a novel framework that probes for hidden retained knowledge in unlearned LLMs through adversarial suffix prompting. LURK automatically generates adversarial prompt suffixes designed to elicit residual knowledge about the Harry Potter domain, a commonly used benchmark for unlearning. Our experiments reveal that even models deemed successfully unlearned can leak idiosyncratic information under targeted adversarial conditions, highlighting critical limitations of current unlearning evaluation standards. By uncovering latent knowledge through indirect probing, LURK offers a more rigorous and diagnostic tool for assessing the robustness of unlearning algorithms. All code will be publicly available.

Training generalist agents capable of adapting to diverse scenarios requires interactive environments for self-exploration. However, interactive environments remain critically scarce, and existing synthesis methods suffer from significant limitations regarding environmental diversity and scalability. To address these challenges, we introduce ScaleEnv, a framework that constructs fully interactive environments and verifiable tasks entirely from scratch. Specifically, ScaleEnv ensures environment reliability through procedural testing, and guarantees task completeness and solvability via tool dependency graph expansion and executable action verification. By enabling agents to learn through exploration within ScaleEnv, we demonstrate significant performance improvements on unseen, multi-turn tool-use benchmarks such as τ^2-Bench and VitaBench, highlighting strong generalization capabilities. Furthermore, we investigate the relationship between increasing number of domains and model generalization performance, providing empirical evidence that scaling environmental diversity is critical for robust agent learning.

LLM-based agents are assumed to integrate environmental observations into their reasoning: discovering highly relevant but unexpected information should naturally lead to a model exploiting its own discoveries. We show that this assumption is false for current LLM-based agents, which struggle to reflect or react to unexpected information. Across three benchmarks (Terminal-Bench, SWE-Bench, AppWorld), we inject complete task solutions into the agent environments to deliberately expose a task's solution to a model. While agents discover these solutions on Terminal-Bench in 79-81% of runs, they interact, or exploit, them in only 37-50% of cases. This gap is starkest in AppWorld: agents see documentation stating that a command "returns the complete solution to this task" in over 90% of attempts but exploit this in fewer than 7% of trials. We show that agents lack what we call environmental curiosity: the capability to recognize and investigate unexpected but relevant observations in response to environmental stimuli. We identify three main factors influencing environmental curiosity: available tools in the agent scaffold, test-time compute, and training data distribution. Our findings identify configurations that maximize curiosity also achieve the best performance on the unmodified benchmarks. Yet even jointly optimized agents still ignore discovered solutions in the majority of trials: current agents use the environment to fetch expected information, but not to revise their strategy or maximally exploit useful stimuli.

Coding agents are increasingly used to build programs from scratch as software engineers do rather than to resolve issues within a repository or to generate a repository from a given requirements document. However, scalable verifiable training environments that force agents to discover what is required and to decide how the system should be designed remain scarce. Moreover, existing agent workflows also cannot sustain long-horizon development in which the agent explores the reference program before implementation. To address these challenges, we introduce ProgramGym, a framework consisting of Program-Env and Program-Agent for scalable training environments of long-horizon program reproduction agents. Specifically, Program-Env constructs verifiable training environments from repositories through a feedback-driven loop that repairs runtime sandbox builds and the quality of black-box tests. Using this novel pipeline, we obtain 1,648 training environments spanning C, C++, Go, and Rust. Program-Agent combines a role-isolated workflow of analysis, implementation, and review with revision management for compilation checks and failure recovery. Experiments on ProgramBench show that both Program-Env and Program-Agent improve performance on program reproduction. Fine-tuning on trajectories with these constructed environments raises the performance of Qwen3.8-27B from 34.87% to 42.84%. Program-Agent further raises the base and SFT models to 55.71% and 56.25%, both exceeding some flagship models. We will release the ProgramGym to support further research on coding agents that build programs from scratch.

As AI agents begin to automate AI R&D, we need ways to assess whether their outputs are safe to deploy, even when the agents themselves may be untrusted. AI control offers one such approach: rather than trusting the agent, it treats it as a potential adversary and uses a monitor to detect covert sabotage before deployment. We evaluate AI control for automated AI R&D with ResearchArena, a framework spanning four long-horizon tasks: safety post-training, capabilities post-training, CUDA-kernel optimization, and inference-server optimization. Because the deliverable in AI R&D is an artifact that will be deployed and run, we pair each main task with two kinds of hidden side task: an embedded side task that sabotages the submitted model, adapter, kernel, or server itself, and an independent side task that takes a covert out-of-scope action in the sandbox while leaving the artifact honest. We evaluate frontier agents at both sabotage and monitoring, across four monitor types that vary along two axes: whether the monitor may execute and probe the produced artifact, and whether it sees the agent's chain-of-thought. We find that sabotage hidden in the training data is the hardest to catch, flagged fewer than half the time. Letting the monitor run experiments on the artifact, rather than only read the trajectory, helps, but it is not enough: monitors still miss embedded sabotage by inspecting only the surface, by noticing the anomaly but explaining it away, or by probing the artifact with the wrong test. We release ResearchArena as a modular framework for evaluating sabotage and control in automated AI R&D.

Machine unlearning is an emerging technique that removes the influence of a subset of training data (forget set) from a model without full retraining, with applications including privacy protection, content moderation, and model correction. The key challenge lies in ensuring that the model completely forgets the knowledge of the forget set without compromising its overall utility. Existing unlearning methods for large language models (LLMs) often utilize auxiliary language models, retain datasets, or even commercial AI services for effective unlearning and maintaining the model utility. However, dependence on these external resources is often impractical and could potentially introduce additional privacy risks. In this work, we propose direct token optimization (DTO), a novel self-contained unlearning approach for LLMs that directly optimizes the token level objectives and eliminates the need for external resources. Given a sequence to unlearn, we identify two categories of tokens: target tokens, which capture critical knowledge for unlearning, and the remaining non-target tokens, which are crucial for maintaining the model utility. The former are used to optimize the unlearning objective, while the latter serve to preserve the model's performance. The experimental results show that the proposed DTO achieves up to 16.8times improvement in forget quality on several benchmark datasets than the latest baselines while maintaining a comparable level of model utility.

This paper investigates the faithfulness of multimodal large language model (MLLM) agents in the graphical user interface (GUI) environment, aiming to address the research question of whether multimodal GUI agents can be distracted by environmental context. A general setting is proposed where both the user and the agent are benign, and the environment, while not malicious, contains unrelated content. A wide range of MLLMs are evaluated as GUI agents using our simulated dataset, following three working patterns with different levels of perception. Experimental results reveal that even the most powerful models, whether generalist agents or specialist GUI agents, are susceptible to distractions. While recent studies predominantly focus on the helpfulness (i.e., action accuracy) of multimodal agents, our findings indicate that these agents are prone to environmental distractions, resulting in unfaithful behaviors. Furthermore, we switch to the adversarial perspective and implement environment injection, demonstrating that such unfaithfulness can be exploited, leading to unexpected risks.

Experimental materials-science labs routinely accumulate synthesis and characterization data across spreadsheets, instrument export files, and ad-hoc folders, with no durable link between a sample's recipe, its measured properties, and whether a similar composition has been reported elsewhere. We present Stage 1 of an open-source data infrastructure closing this loop for a single working alloy-synthesis lab: a relational schema for composition and synthesis records, a strict composition/mass calculator, three empirical alloy-design screening metrics (VEC, atomic-size mismatch, mixing enthalpy) plus a melt/boil-point synthesis-feasibility check, an automated cross-referencing layer against three public materials databases (Materials Project, OQMD, Alexandria) with deduplication and adjustable relevance filtering, and a desktop GUI unifying the workflow behind a single "new entry" action. Built and used by a practicing researcher rather than a software team, we report it as a systems/methodology contribution - the architecture, the concrete engineering failures encountered while building it, and how each was diagnosed and fixed - rather than new materials-science results. We argue the specific defects found - a parser ambiguity that produced silent wrong answers, a partial code edit that broke a dependency chain invisibly, a database-metadata bug traceable to an un-cast numeric type, and a credential-hygiene gap that .gitignore alone did not close - generalize to any small scientific-software project built by a domain expert without formal software-engineering training. Both the system and the lessons from building it are reported here; two follow-on papers cover the validated characterization pipelines (Stage 2) and the retrieval-augmented synthesis advisor and CALPHAD thermodynamics module (Stage 3) built on this foundation.

In Open Source Software, the source code and any other resources available in a project can be viewed or reused by anyone subject to often permissive licensing restrictions. In contrast to some studies of dependency-based reuse supported via package managers, no studies of OSS-wide copy-based reuse exist. This dataset seeks to encourage the studies of OSS-wide copy-based reuse by providing copying activity data that captures whole-file reuse in nearly all OSS. To accomplish that, we develop approaches to detect copy-based reuse by developing an efficient algorithm that exploits World of Code infrastructure: a curated and cross referenced collection of nearly all open source repositories. We expect this data to enable future research and tool development that support such reuse and minimize associated risks.

Graphical User Interface (GUI) agents are increasingly deployed to interact with online web services, yet their exposure to open-world content renders them vulnerable to Environmental Injection Attacks (EIAs). In these attacks, an attacker can inject crafted triggers into website to manipulate the behavior of GUI agents used by other users. In this paper, we find that most existing EIA studies fall short of realism. In particular, they fail to capture the dynamic nature of real-world web content, often assuming that a trigger's on-screen position and surrounding visual context remain largely consistent between training and testing. To better reflect practice, we introduce a realistic dynamic-environment threat model in which the attacker is a regular user and the trigger is embedded within a dynamically changing environment. Under this threat model, existing approaches largely fail, suggesting that their effectiveness in exposing GUI agent vulnerabilities has been substantially overestimated. To expose the hidden vulnerabilities of existing GUI agents effectively, we propose Chameleon, an attack framework with two key novelties designed for dynamic environments. (1) To synthesize more realistic training data, we introduce LLM-Driven Environment Simulation, which automatically generates diverse, high-fidelity webpage simulations that mimic the variability of real-world dynamic environments. (2) To optimize the trigger more effectively, we introduce Attention Black Hole, which converts attention weights into explicit supervisory signals. This mechanism encourages the agent to remain insensitive to irrelevant surrounding content, thereby improving robustness in dynamic environments. We evaluate Chameleon on six realistic websites and four representative LVLM-powered GUI agents, where it significantly outperforms existing methods.

The field of artificial intelligence (AI) has experienced remarkable progress in recent years, driven by the widespread adoption of open-source machine learning models in both research and industry. Considering the resource-intensive nature of training on vast datasets, many applications opt for models that have already been trained. Hence, a small number of key players undertake the responsibility of training and publicly releasing large pre-trained models, providing a crucial foundation for a wide range of applications. However, the adoption of these open-source models carries inherent privacy and security risks that are often overlooked. To provide a concrete example, an inconspicuous model may conceal hidden functionalities that, when triggered by specific input patterns, can manipulate the behavior of the system, such as instructing self-driving cars to ignore the presence of other vehicles. The implications of successful privacy and security attacks encompass a broad spectrum, ranging from relatively minor damage like service interruptions to highly alarming scenarios, including physical harm or the exposure of sensitive user data. In this work, we present a comprehensive overview of common privacy and security threats associated with the use of open-source models. By raising awareness of these dangers, we strive to promote the responsible and secure use of AI systems.

Large language model (LLM) agents integrated with external tools are vulnerable to indirect prompt injections embedded in environmental states. However, existing studies largely rely on manually implemented or reused environments, stochastic LLM-based tool simulation, and predefined injection locations, limiting scalable security research across broader domains. To bridge this gap, we propose **ToolHazard**, a scalable adversarial environment synthesis framework that reduces human engineering and supports expansion with additional seed domains and compute. Through an Environment Simulator, an Attacker Agent, and a User Simulator, ToolHazard synthesizes executable stateful environments, discovers viable injection points and generates environment-specific payloads, and constructs state-grounded long-horizon tasks. Based on ToolHazard, we build **ToolHazard-Bench** for stress-testing agents under complex workflows and diverse environmental attacks. Experiments reveal substantial agent vulnerabilities and show that injection timing and placement affect attack effectiveness. Moreover, ToolHazard-generated alignment data improves security on both ToolHazard-Bench and AgentDojo while preserving benign task utility.

Refusal on harmful prompts is a key safety behaviour in instruction-tuned large language models (LLMs), yet the internal causes of this behaviour remain poorly understood. We study two public instruction-tuned models, Gemma-2-2B-IT and LLaMA-3.1-8B-IT, using sparse autoencoders (SAEs) trained on residual-stream activations. Given a harmful prompt, we search the SAE latent space for feature sets whose ablation flips the model from refusal to compliance, demonstrating causal influence and creating a jailbreak. Our search proceeds in three stages: (1) Refusal Direction: find a refusal-mediating direction and collect SAE features near that direction; (2) Greedy Filtering: prune to a minimal set; and (3) Interaction Discovery: fit a factorization machine (FM) that captures nonlinear interactions among the remaining active features and the minimal set. This pipeline yields a broad set of jailbreak-critical features, offering insight into the mechanistic basis of refusal. Moreover, we find evidence of redundant features that remain dormant unless earlier features are suppressed. Our findings highlight the potential for fine-grained auditing and targeted intervention in safety behaviours by manipulating the interpretable latent space.

Application Programming Interfaces (APIs) facilitate the integration of third-party dependencies within the code of client applications. However, changes to an API, such as deprecation, modification of parameter names or types, or complete replacement with a new API, can break existing client code. These changes are called breaking dependency updates; It is often tedious for API users to identify the cause of these breaks and update their code accordingly. In this paper, we explore the use of Large Language Models (LLMs) to automate client code updates in response to breaking dependency updates. We evaluate our approach on the BUMP dataset, a benchmark for breaking dependency updates in Java projects. Our approach leverages LLMs with advanced prompts, including information from the build process and from the breaking dependency analysis. We assess effectiveness at three granularity levels: at the build level, the file level, and the individual compilation error level. We experiment with five LLMs: Google Gemini-2.0 Flash, OpenAI GPT4o-mini, OpenAI o3-mini, Alibaba Qwen2.5-32b-instruct, and DeepSeek V3. Our results show that LLMs can automatically repair breaking updates. Among the considered models, OpenAI's o3-mini is the best, able to completely fix 27% of the builds when using prompts that include contextual information such as the buggy line, API differences, error messages, and step-by-step reasoning instructions. Also, it fixes 78% of the individual compilation errors. Overall, our findings demonstrate the potential for LLMs to fix compilation errors due to breaking dependency updates, supporting developers in their efforts to stay up-to-date with changes in their dependencies.

Deep Neural Networks (DNNs) are being adopted as components in software systems. Creating and specializing DNNs from scratch has grown increasingly difficult as state-of-the-art architectures grow more complex. Following the path of traditional software engineering, machine learning engineers have begun to reuse large-scale pre-trained models (PTMs) and fine-tune these models for downstream tasks. Prior works have studied reuse practices for traditional software packages to guide software engineers towards better package maintenance and dependency management. We lack a similar foundation of knowledge to guide behaviors in pre-trained model ecosystems. In this work, we present the first empirical investigation of PTM reuse. We interviewed 12 practitioners from the most popular PTM ecosystem, Hugging Face, to learn the practices and challenges of PTM reuse. From this data, we model the decision-making process for PTM reuse. Based on the identified practices, we describe useful attributes for model reuse, including provenance, reproducibility, and portability. Three challenges for PTM reuse are missing attributes, discrepancies between claimed and actual performance, and model risks. We substantiate these identified challenges with systematic measurements in the Hugging Face ecosystem. Our work informs future directions on optimizing deep learning ecosystems by automated measuring useful attributes and potential attacks, and envision future research on infrastructure and standardization for model registries.

Public repositories host millions of fine-tuned models, yet community usage remains disproportionately concentrated on a small number of foundation checkpoints. We investigate whether this concentration reflects efficient market selection or if superior models are systematically overlooked. Through an extensive evaluation of over 2,000 models, we show the prevalence of "hidden gems", unpopular fine-tunes that significantly outperform their popular counterparts. Notably, within the Llama-3.1-8B family, we find rarely downloaded checkpoints that improve math performance from 83.2% to 96.0% without increasing inference costs. However, discovering these models through exhaustive evaluation of every uploaded model is computationally infeasible. We therefore formulate model discovery as a Multi-Armed Bandit problem and accelerate the Sequential Halving search algorithm by using shared query sets and aggressive elimination schedules. Our method retrieves top models with as few as 50 queries per candidate, accelerating discovery by over 50x.

Large language models (LLMs) extended as systems, such as ChatGPT, have begun supporting third-party applications. These LLM apps leverage the de facto natural language-based automated execution paradigm of LLMs: that is, apps and their interactions are defined in natural language, provided access to user data, and allowed to freely interact with each other and the system. These LLM app ecosystems resemble the settings of earlier computing platforms, where there was insufficient isolation between apps and the system. Because third-party apps may not be trustworthy, and exacerbated by the imprecision of natural language interfaces, the current designs pose security and privacy risks for users. In this paper, we evaluate whether these issues can be addressed through execution isolation and what that isolation might look like in the context of LLM-based systems, where there are arbitrary natural language-based interactions between system components, between LLM and apps, and between apps. To that end, we propose IsolateGPT, a design architecture that demonstrates the feasibility of execution isolation and provides a blueprint for implementing isolation, in LLM-based systems. We evaluate IsolateGPT against a number of attacks and demonstrate that it protects against many security, privacy, and safety issues that exist in non-isolated LLM-based systems, without any loss of functionality. The performance overhead incurred by IsolateGPT to improve security is under 30% for three-quarters of tested queries.

As Large Language Models (LLMs) are increasingly deployed as autonomous agents in complex and long horizon settings, it is critical to evaluate their ability to sabotage users by pursuing hidden objectives. We study the ability of frontier LLMs to evade monitoring and achieve harmful hidden goals while completing a wide array of realistic tasks. We evaluate a broad range of frontier LLMs using SHADE (Subtle Harmful Agent Detection & Evaluation)-Arena, the first highly diverse agent evaluation dataset for sabotage and monitoring capabilities of LLM agents. SHADE-Arena consists of complex pairs of benign main tasks and harmful side objectives in complicated environments. Agents are evaluated on their ability to complete the side task without appearing suspicious to an LLM monitor. When measuring agent ability to (a) complete the main task, (b) complete the side task, and (c) avoid detection, we find that the best performing frontier models score 27% (Claude 3.7 Sonnet) and 15% (Gemini 2.5 Pro) as sabotage agents when overseen by Claude 3.6 Sonnet. For current frontier models, success on the side task relies heavily on having access to a hidden scratchpad that is not visible to the monitor. We also use SHADE-Arena to measure models' monitoring abilities, with the top monitor (Gemini 2.5 Pro) achieving an AUC of 0.87 at distinguishing benign and malign transcripts. We find that for now, models still struggle at sabotage due to failures in long-context main task execution. However, our measurements already demonstrate the difficulty of monitoring for subtle sabotage attempts, which we expect to only increase in the face of more complex and longer-horizon tasks.

Before developers can use a software project, they must first get it to run. This often requires installing the right language runtime, system packages, build tools, and test commands. Recent LLM-based agents automate environment setup by inspecting a repository, writing setup scripts, running them, and repairing the resulting errors. However, these agents still follow a sequential trial-and-error loop: they try one setup route, wait for slow commands to finish, observe the failure, and only then decide what to try next. Because environment commands often take long to run, this sequential workflow makes setup slow and limits exploration to one route at a time. We present SpecBuild, an LLM-based environment setup system designed to be both faster and more successful through speculative forking. SpecBuild explores candidate setup branches in parallel, returns the first committed environment, and discards the rest. It applies this design at two granularities: strategy-level forking explores alternative setup routes, such as different base images and dependency policies, while command-level forking predicts failures of running commands and prepares repair branches before those failures are observed. We evaluate SpecBuild on Multi-Docker-Eval, covering 334 tasks from 40 repositories and 9 programming languages. Across three backbone models, SpecBuild improves the success rate by 6.4 to 12.0 percentage points over the strongest baseline, while reducing average setup time by 22.4% to 41.6%. These results show that LLM-based environment setup need not trade speed for reliability: parallel speculative exploration can make agents both faster and more successful.

As multi-agent systems enter high-stakes domains, the possibility that agents may circumvent safety boundaries is a growing concern. Prior work has examined this risk primarily in adversarial settings, where agents are instructed or rewarded to communicate covertly and evade oversight. We show that benign agents can cross the same boundaries without adversarial incentives. We emulate a software-engineering workflow in which a planner represents a company hiring an external developer. The planner writes requirements and holds a company credential it is instructed not to disclose to the developer; a monitor screens their exchanges. Seven of nine tested frontier models disguise the credential in their requirements to help the developer recover it while evading the monitor, even after completing their assigned objective. For example, across 6,000 episodes with DeepSeek-V4-Pro, the planner attempts concealment in 16.9%; in 0.9%, the credential evades the monitor and is recovered and used by the developer. Despite the low per-episode rate, the risk compounds across repeated exchanges: at 0.9%, 105 independent episodes carry a 61.3% chance of at least one breach. Our analysis suggests that the planner misinterprets disguised disclosure as helping the developer and reads the nondisclosure rule as banning plaintext, not character codes or riddles. We call this behavior covert assistance. Monitoring favors concealment, while concealment gives way to direct disclosure when the recipient is presented as a human rather than an AI agent, so leakage persists. Giving the monitor the credential improves detection but requires trusting it with the secret. These risks, in models already used for software engineering, challenge oversight to distinguish authorized cooperation from task-advancing assistance that crosses safety boundaries.

AI agents are increasingly being deployed to automate tasks, often based on underspecified user instructions. Making unwarranted assumptions to compensate for the missing information and failing to ask clarifying questions can lead to suboptimal outcomes, safety risks due to tool misuse, and wasted computational resources. In this work, we study the ability of LLM agents to handle underspecified instructions in interactive code generation settings by evaluating proprietary and open-weight models on their performance across three key steps: (a) detecting underspecificity, (b) asking targeted clarification questions, and (c) leveraging the interaction to improve performance in underspecified scenarios. We introduce Ambig-SWE, an underspecified variant of SWE-Bench Verified, specifically designed to evaluate agent behavior under ambiguity and interaction. Our findings reveal that models struggle to distinguish between well-specified and underspecified instructions. However, when models interact for underspecified inputs, they effectively obtain vital information from the user leading to significant improvements in performance, up to 74% over the non-interactive settings, underscoring the value of effective interaction. Our study highlights critical gaps in how current state-of-the-art models handle missing information in complex software engineering tasks and structures the evaluation into distinct steps to enable targeted improvements.

Note: This paper describes an older version of DeepLIFT. See https://arxiv.org/abs/1704.02685 for the newer version. Original abstract follows: The purported "black box" nature of neural networks is a barrier to adoption in applications where interpretability is essential. Here we present DeepLIFT (Learning Important FeaTures), an efficient and effective method for computing importance scores in a neural network. DeepLIFT compares the activation of each neuron to its 'reference activation' and assigns contribution scores according to the difference. We apply DeepLIFT to models trained on natural images and genomic data, and show significant advantages over gradient-based methods.

Stealth, the discipline of achieving an objective without revealing your presence, capabilities, or collected intelligence, is what separates sophisticated operators from detectable ones. Elite security researchers and advanced persistent threats achieve their objectives unnoticed; autonomous agents increasingly inherit the same offensive tasks, but do they inherit the tradecraft? We introduce StealthBench,a benchmark that measures operational stealth in autonomous offensive-security agents across six operational security (OPSEC) dimensions. We extract 11 hand-verified OPSEC incidents from real bug-bounty and red-team trajectories, expanded into 14 dockerized task scenarios, where agents, despite finding real vulnerabilities, committed stealth failures inconsistent with standard operational tradecraft: embedding credentials in public uploads, deleting production resources to prove access, force-adding uninvolved users to demonstrate a race condition. We evaluate agent trajectories using a 3-model large language model (LLM) judge panel with majority-vote aggregation, measuring safe success rate (solved and stealthy), Stealth@Solve (tradecraft quality among successful solves), and reckless solve rate (solved but cover blown). Our results show that no model exceeds 54% safe success rate (the compound metric requiring both task completion and stealth), confirming that OPSEC failures are systematic across model families. We release StealthBench as a public benchmark to support both the development of stealth-aware agents and automated OPSEC monitoring for autonomous offensive-security deployments. The interactive leaderboard, evaluation harness, and dataset are available at https://stealthbench.com.

In the rapidly evolving software development landscape, Python stands out for its simplicity, versatility, and extensive ecosystem. Python packages, as units of organization, reusability, and distribution, have become a pressing concern, highlighted by the considerable number of vulnerability reports. As a scripting language, Python often cooperates with other languages for performance or interoperability. This adds complexity to the vulnerabilities inherent to Python packages, and the effectiveness of current vulnerability detection tools remains underexplored. This paper addresses these gaps by introducing PyVul, the first comprehensive benchmark suite of Python-package vulnerabilities. PyVul includes 1,157 publicly reported, developer-verified vulnerabilities, each linked to its affected packages. To accommodate diverse detection techniques, it provides annotations at both commit and function levels. An LLM-assisted data cleansing method is incorporated to improve label accuracy, achieving 100% commit-level and 94% function-level accuracy, establishing PyVul as the most precise large-scale Python vulnerability benchmark. We further carry out a distribution analysis of PyVul, which demonstrates that vulnerabilities in Python packages involve multiple programming languages and exhibit a wide variety of types. Moreover, our analysis reveals that multi-lingual Python packages are potentially more susceptible to vulnerabilities. Evaluation of state-of-the-art detectors using this benchmark reveals a significant discrepancy between the capabilities of existing tools and the demands of effectively identifying real-world security issues in Python packages. Additionally, we conduct an empirical review of the top-ranked CWEs observed in Python packages, to diagnose the fine-grained limitations of current detection tools and highlight the necessity for future advancements in the field.

Large language models are increasingly deployed as autonomous agents serving users on behalf of companies, placing them in settings where user and deployer interests can conflict. When an agent knows that a user is owed something its deployer would prefer to deny, does it remain honest? Answering this is difficult because false statements can reflect either ignorance or hallucination rather than deception. To address this challenge, we introduce KnownLieBench , a knowledge-verified benchmark that first confirms through a neutral probe that an agent knows a user's entitlement, and then evaluates whether it makes false claims once an incentive to deny that entitlement is introduced. Specifically, KnownLieBench covers eight customer-service domains and 112 grounded cases, conducts multi-round dialogues with a trust-tracking customer agent, and separates deception emerging from incentive alone from deception produced under explicit instruction. Across eighteen proprietary and open-weight models, emergent deception varies substantially across model families and domains. We further use the benchmark for post-training, finding that honesty-directed fine-tuning reduces deception under incentive, while deception-graded fine-tuning increases lie success on honest-control dialogues without increasing lie frequency under incentive. By verifying entitlement knowledge before scoring deceptive behavior, KnownLieBench reduces the confound between lying and not knowing and enables more rigorous auditing and steering of agent honesty.

Package confusion attacks such as typosquatting threaten software supply chains. Attackers make packages with names that syntactically or semantically resemble legitimate ones, tricking engineers into installing malware. While prior work has developed defenses against package confusions in some software package registries, notably NPM, PyPI, and RubyGems, gaps remain: high false-positive rates; generalization to more software package ecosystems; and insights from real-world deployment. In this work, we introduce ConfuGuard, a solution designed to address the challenges posed by package confusion threats. We begin by presenting the first empirical analysis of benign signals derived from prior package confusion data, uncovering their threat patterns, engineering practices, and measurable attributes. We observed that 13.3% of real package confusion attacks are initially stealthy, so we take that into consideration and refined the definitions. Building on state-of-the-art approaches, we extend support from three to six software package registries, and leverage package metadata to distinguish benign packages. Our approach significantly reduces 64% false-positive (from 77% to 13%), with acceptable additional overhead to filter out benign packages by analyzing the package metadata. ConfuGuard is in production at our industry partner, whose analysts have already confirmed 301 packages detected by ConfuGuard as real attacks. We share lessons learned from production and provide insights to researchers.

We discover that DALLE-2 seems to have a hidden vocabulary that can be used to generate images with absurd prompts. For example, it seems that Apoploe vesrreaitais means birds and Contarra ccetnxniams luryca tanniounons (sometimes) means bugs or pests. We find that these prompts are often consistent in isolation but also sometimes in combinations. We present our black-box method to discover words that seem random but have some correspondence to visual concepts. This creates important security and interpretability challenges.

LLM agents span command-line interfaces (e.g., Codex) and agent frameworks (e.g., LangChain), integrating backend LLMs with harness code that parses model outputs, controls agent loops, and manages context. Both the harness and LLM-generated responses jointly shape an agent's execution. This architecture gives rise to bugs that cannot be readily understood by inspecting either component alone, because some bugs occur only when a particular LLM response elicits an abnormal reaction from the agent. Prior empirical studies of agent bugs have largely attributed failures either to limited model capabilities or to harness-side defects, such as outdated APIs and configuration misalignment, without characterizing these AR bugs. We conduct the first empirical study focused on agent-reactive (AR) bugs. Through manual analysis of 255 bug reports from Codex, Gemini-CLI, LangChain, and CrewAI, we construct a two-axis taxonomy covering observable symptoms and the LLM behaviors that trigger them. Our findings show that many AR bugs manifest as silent errors without well-defined test oracles, which makes detection difficult. The stochasticity of LLM responses further complicates bug reproduction. We additionally examine fixes proposed by users and implemented by developers. This analysis exposes a mismatch: users frequently advocate harness-side guardrails, whereas developers may attribute the issue to the LLM or respond slowly to user-proposed fixes. These findings point to the need for mechanisms that help users and developers understand the root causes and resolutions of AR bugs. Overall, the study highlights challenges specific to LLM agents and motivates the design of test oracles, reproduction support, and fault-localization techniques for AR bugs.

Open-weight LLMs are released not only as fixed products but also as substrates for downstream fine-tuning. This openness, however, creates legal and ethical risks because users may misuse fine-tuning to instill illicit knowledge or enable hostile operations. Model providers therefore need apre-release defense against such acquisition, motivating the problem of preemptive unlearning. Unlike retrospective unlearning, which removes capabilities already present in a fixed model, preemptive unlearning seeks to prevent their acquisition under unseen attack data and future fine-tuning procedures. Despite its practical importance, this setting remains largely unexplored, presents distinct challenges, and is therefore the central focus of our work. We first verify that existing retrospective methods provide insufficient pre-release protection. Even when forbidden capabilities are suppressed in current outputs, forbidden-domain data can still induce gradients through internal pathways, enabling later acquisition. Motivated by this finding, we propose a gradient-sealing principle that blocks these pathways by pushing relevant pre-activations into the negative region, where ReLU-family activations exhibit zero or near-zero derivatives. Experiments across multiple LLM families demonstrate our stronger resistance to downstream acquisition than retrospective baselines, validating gradient sealing as an effective mechanism for pre-release protection.

The proliferation of Large Language Models (LLMs) accessed via black-box APIs introduces a significant trust challenge: users pay for services based on advertised model capabilities (e.g., size, performance), but providers may covertly substitute the specified model with a cheaper, lower-quality alternative to reduce operational costs. This lack of transparency undermines fairness, erodes trust, and complicates reliable benchmarking. Detecting such substitutions is difficult due to the black-box nature, typically limiting interaction to input-output queries. This paper formalizes the problem of model substitution detection in LLM APIs. We systematically evaluate existing verification techniques, including output-based statistical tests, benchmark evaluations, and log probability analysis, under various realistic attack scenarios like model quantization, randomized substitution, and benchmark evasion. Our findings reveal the limitations of methods relying solely on text outputs, especially against subtle or adaptive attacks. While log probability analysis offers stronger guarantees when available, its accessibility is often limited. We conclude by discussing the potential of hardware-based solutions like Trusted Execution Environments (TEEs) as a pathway towards provable model integrity, highlighting the trade-offs between security, performance, and provider adoption. Code is available at https://github.com/sunblaze-ucb/llm-api-audit

Training terminal agents requires scalable executable supervision, yet synthesizing high-quality terminal tasks remains challenging. Each task couples an instruction, an initialized environment, a reference solution, and an executable verifier; if these artifacts are generated from inconsistent assumptions, the resulting task may be unsolvable or incorrectly evaluated. Meanwhile, multi-stage synthesis can discard the goals, dependencies, state transitions, and procedural constraints encoded in the original sources. We present FACET (Fine-grained Agentic Construction of Executable Tasks), a framework that addresses both information preservation and cross-artifact consistency. FACET reconstructs related agent skills into coherent, information-rich scenarios, then realizes and repairs the execution environment before generating the final task artifacts. The resulting container state serves as shared grounding for the instruction, solution, and verifier, while execution-based validation and targeted repair correct artifact-specific failures without unnecessarily regenerating valid components. FACET produces complex terminal tasks with dense executable checks, and successful trajectories collected from these tasks provide effective, data-efficient supervision. Fine-tuning models across multiple scales consistently improves performance on Terminal-Bench 2.1, while analyses of alternative generation schemes support the importance of environment-grounded construction for task validity and solution-verifier alignment. These results establish source-intent preservation and shared executable-state grounding as key principles for scalable terminal-task synthesis.

This study reveals a previously unexplored vulnerability in the safety alignment of Large Language Models (LLMs). Existing aligned LLMs predominantly respond to unsafe queries with refusals, which often begin with a fixed set of prefixes (I'm sorry). We demonstrate that this rigid refusal pattern is a vulnerability and introduce a novel refusal unlearning technique that exploits it. Specifically, we fine-tune LLMs using merely 1,000 benign samples, where each response is prepended with a refusal prefix. The underlying intuition is to disrupt the refusal completion pathway, thereby driving the model to forget how to refuse while following harmful instructions. This intuition is further supported by theoretical proofs. We apply this approach to a total of 16 LLMs, including various open-source models from Llama, Qwen, and Gemma families, as well as closed-source models such as Gemini and GPT. Experimental results show that the safety scores of previously aligned LLMs degrade both consistently and substantially. Importantly, we verify that the observed gain cannot be attributed to plain fine-tuning or random prefix effects. Our findings suggest that current safety alignment may rely heavily on token sequence memorization rather than reasoning, motivating future work beyond simple refusal mechanisms. Code has been released: https://github.com/guoyang9/refusal-unlearning.

To operate robustly in open-world environments, autonomous agents should be able to infer the behavior of unfamiliar systems through interaction alone, even in the absence of documentation. However, existing tool-use benchmarks expose semantic tool schemas in static environments, allowing agents to rely on prior knowledge rather than autonomous discovery. To address this limitation, we introduce ScrambleToolBench, an interactive terminal benchmark designed to isolate behavioral reasoning. By removing semantic cues and enforcing a continuous task curriculum, the benchmark requires agents to uncover hidden tool behaviors entirely through trial-and-error interaction. The benchmark further introduces dynamic challenges, including mapping drift, stochastic action failures, and temporal execution windows, to evaluate whether agents can revise and adapt their hypotheses as the environment changes. Our evaluation of state-of-the-art language models reveals that successful initial discovery does not translate into robust adaptation. When faced with structural changes such as mapping drift, agents fail to use deductive strategies such as cycle tracing, and instead exhibit belief inertia or fall back to exhaustive search. Increasing test-time reasoning only amplifies this expensive brute-force search rather than enabling deductive recovery. While equipping agents with persistent memory reduces compounding errors, they remain unable to efficiently infer structural changes, highlighting a gap in current agent reasoning.

Command-line interfaces (CLIs) provide structured textual environments for system administration. Explorations have been performed using pre-trained language models (PLMs) to simulate these environments for safe interaction in high-risk environments. However, their use has been constrained to frozen, large parameter models like GPT. For smaller architectures to reach a similar level of believability, a rich dataset of CLI interactions is required. Existing public datasets focus on mapping natural-language tasks to commands, omitting crucial execution data such as exit codes, outputs, and environmental side effects, limiting their usability for behavioral modeling. We introduce a Shell Input -Output Environment (ShIOEnv), which casts command construction as a Markov Decision Process whose state is the partially built sequence and whose actions append arguments. After each action, ShIOEnv executes the candidate and returns its exit status, output, and progress toward a minimal-length behavioral objective. Due to the intractable nature of the combinatorial argument state-action space, we derive a context-free grammar from man pages to mask invalid arguments from being emitted. We explore random and proximal-policy optimization (PPO)-optimized sampling of unrestricted and grammar-masked action spaces to produce four exploration strategies. We observed that grammar masking and PPO significantly improve sample efficiency to produce a higher quality dataset (maximizing the number of arguments while minimizing redundancies). Policy-generated datasets of shell input-output behavior pairs are used to fine-tune CodeT5, where we observe 85% improvements in BLEU-4 when constraining the action space to grammar productions with an additional 26% improvement when applying PPO. The ShIOEnv environment and curated command behavior datasets are released for use in future research.

The npm (Node Package Manager) ecosystem is the most important package manager for JavaScript development with millions of users. Consequently, a plethora of earlier work investigated how vulnerability reporting, patch propagation, and in general detection as well as resolution of security issues in such ecosystems can be facilitated. However, understanding the ground reality of security-related issue reporting by users (and bots) in npm-along with the associated challenges has been relatively less explored at scale. In this work, we bridge this gap by collecting 10,907,467 issues reported across GitHub repositories of 45,466 diverse npm packages. We found that the tags associated with these issues indicate the existence of only 0.13% security-related issues. However, our approach of manual analysis followed by developing high accuracy machine learning models identify 1,617,738 security-related issues which are not tagged as security-related (14.8% of all issues) as well as 4,461,934 comments made on these issues. We found that the bots which are in wide use today might not be sufficient for either detecting or offering assistance. Furthermore, our analysis of user-developer interaction data hints that many user-reported security issues might not be addressed by developers-they are not tagged as security-related issues and might be closed without valid justification. Consequently, a correlation analysis hints that the developers quickly handle security issues with known solutions (e.g., corresponding to CVE). However, security issues without such known solutions (even with reproducible code) might not be resolved. Our findings offer actionable insights for improving security management in open-source ecosystems, highlighting the need for smarter tools and better collaboration. The data and code for this work is available at https://doi.org/10.5281/zenodo.15614029

Agent environments are essential for large language model (LLM) development. They use programmatic verifiers to turn complex outcomes into rewards. Verifiers can be attacked into awarding full credit to cheating agents without completing the intended task, causing inaccurate evaluation and misalignment in training. As agent environments scale in numbers and difficulty, so must the repairing of exploitable verifiers. To this end, we ask whether today's model with scaled compute can automatically harden agent environments to prevent potentially stronger future models from reward hacking. We introduce a scalable, fully automatic approach to weak-to-strong verifier hardening. It first sets up an isolated verifier sandbox that agents cannot directly modify, then uses a hacker-fixer loop to find and repair remaining verifier vulnerabilities. We apply this loop to Terminal Bench, KernelBench, and SETA, substantially reducing attack success across these agent environments. The iterative design scales with the hardening budget, producing progressively stronger defenses as the number of loop iterations increases and outperforming the static, non-iterative baseline. An earlier version of our hacker-fixer loop has been adopted by the official Terminal Bench repository as a way to audit exploits in its environments. Project page: https://verifier-hardening.github.io/.

Tool-augmented LLM agents can harbor implicit state that persists across sessions, activates through events, and propagates across agent boundaries---largely invisible to standard debugging. We formalize this as Persistent Semantic Entities (PSEs): constructs defined by name binding, event triggering, and cross-boundary propagation, and evaluate them across 24 models from 11 families (1.5B--1T parameters). First, every tested model is susceptible (20--100% on the 20-model susceptibility panel), with name binding as the necessary and dominant mechanism: without it, contamination is 0%. Second, persistence depends on contamination type rather than scale or deployment: preference contamination persists undecayed on every model probed (100% at t=10) and instruction contamination persists wherever adopted, persona-style injection decays partially (90%to10%), while factual injection is model-dependent---self-corrected on Llama-3.1-8B and GPT-4o-mini but held at ceiling on both Qwen2.5-coder variants, so we do not claim it self-corrects in general. The preference and instruction results hold across providers in our controlled setting. Third, context-isolated self-verification achieves 20--79% reduction (median 36.5%) without oracle references while keyword-based detection produces systematic false positives, and contamination compounds 1.9times along a four-stage agent pipeline (40%to75%). Preference and instruction contamination---persistent, lacking self-correction, and poorly captured by standard monitoring---represent a particularly concerning attack surface for deployed agent systems.

Security experts reverse engineer (decompile) binary code to identify critical security vulnerabilities. The limited access to source code in vital systems - such as firmware, drivers, and proprietary software used in Critical Infrastructures (CI) - makes this analysis even more crucial on the binary level. Even with available source code, a semantic gap persists after compilation between the source and the binary code executed by the processor. This gap may hinder the detection of vulnerabilities in source code. That being said, current research on Large Language Models (LLMs) overlooks the significance of decompiled binaries in this area by focusing solely on source code. In this work, we are the first to empirically uncover the substantial semantic limitations of state-of-the-art LLMs when it comes to analyzing vulnerabilities in decompiled binaries, largely due to the absence of relevant datasets. To bridge the gap, we introduce DeBinVul, a novel decompiled binary code vulnerability dataset. Our dataset is multi-architecture and multi-optimization, focusing on C/C++ due to their wide usage in CI and association with numerous vulnerabilities. Specifically, we curate 150,872 samples of vulnerable and non-vulnerable decompiled binary code for the task of (i) identifying; (ii) classifying; (iii) describing vulnerabilities; and (iv) recovering function names in the domain of decompiled binaries. Subsequently, we fine-tune state-of-the-art LLMs using DeBinVul and report on a performance increase of 19%, 24%, and 21% in the capabilities of CodeLlama, Llama3, and CodeGen2 respectively, in detecting binary code vulnerabilities. Additionally, using DeBinVul, we report a high performance of 80-90% on the vulnerability classification task. Furthermore, we report improved performance in function name recovery and vulnerability description tasks.

Progress in software-engineering agents is increasingly constrained by the scarcity of executable, scalable, and realistic data for training and evaluation. This scarcity stems from three fundamental challenges in existing pipelines: environments are brittle and difficult to reproduce across languages; synthesizing realistic, system-level bugs at scale is computationally expensive; and existing data predominantly consists of short-horizon repairs, failing to capture long-horizon competencies like architectural consistency. We introduce SWE-Hub, an end-to-end system that operationalizes the data factory abstraction by unifying environment automation, scalable synthesis, and diverse task generation into a coherent production stack. At its foundation, the Env Agent establishes a shared execution substrate by automatically converting raw repository snapshots into reproducible, multi-language container environments with standardized interfaces. Built upon this substrate, SWE-Scale engine addresses the need for high-throughput generation, combining cross-language code analysis with cluster-scale validation to synthesize massive volumes of localized bug-fix instances. Bug Agent generates high-fidelity repair tasks by synthesizing system-level regressions involving cross-module dependencies, paired with user-like issue reports that describe observable symptoms rather than root causes. Finally, SWE-Architect expands the task scope from repair to creation by translating natural-language requirements into repository-scale build-a-repo tasks. By integrating these components, SWE-Hub establishes a unified production pipeline capable of continuously delivering executable tasks across the entire software engineering lifecycle.

Long-horizon LLM agents can fail quietly: they settle on one reading of the evidence early, then spend the rest of the run defending it. We call this premature commitment. Final-answer scoring misses the failure mode because it sees only the answer, not whether the process has already collapsed to a stable path. We define representational commitment as cross-run hidden-state convergence at a fixed reasoning step, and use it as an early diagnostic of trajectory consistency. On Llama-3.1-70B running ReAct on HotpotQA, step-4 hidden-state similarity predicts downstream behavioral consistency (r = -0.35, partial r = -0.45), with a localized temporal and layer-wise signature. The signal replicates across Qwen-2.5-72B and Phi-3-14B, and on StrategyQA (r = -0.83). It does not track correctness: committed-wrong and committed-correct questions are not separable in activation similarity. That boundary is central to the claim. Commitment tells us whether an agent has settled, not whether it is right. A runtime monitor detects inconsistent trajectories from hidden states at AUROC up to 0.97 (0.85--0.88 under a stricter split), and a prompting intervention cuts behavioral variance by 28% against a token-matched control while leaving accuracy statistically unchanged. We also test whether the signal can route self-consistency compute; on a harder benchmark it helps only modestly and is matched by a simpler output-based baseline. The result is a diagnostic for a hidden process failure, with clear limits rather than a general accuracy lever.

This paper presents PyResBugs, a curated dataset of residual bugs, i.e., defects that persist undetected during traditional testing but later surface in production, collected from major Python frameworks. Each bug in the dataset is paired with its corresponding fault-free (fixed) version and annotated with multi-level natural language (NL) descriptions. These NL descriptions enable natural language-driven fault injection, offering a novel approach to simulating real-world faults in software systems. By bridging the gap between software fault injection techniques and real-world representativeness, PyResBugs provides researchers with a high-quality resource for advancing AI-driven automated testing in Python systems.

Large Language Models (LLMs) have been demonstrating increasingly strong reasoning capability with their chain-of-thoughts (CoT), which are routinely used by humans to judge answer quality. This reliance creates a powerful yet fragile basis for trust. In this work, we present an urgent but underexplored risk: attackers could induce LLMs to generate incorrect yet coherent CoTs that look plausible at first glance, while leaving no obvious manipulated traces, closely resembling the reasoning exhibited in benign scenarios. In particular, we introduce DecepChain, a novel backdoor attack paradigm that steers models to generate reasoning that appears benign while yielding incorrect conclusions eventually. At a high level, DecepChain exploits LLMs' own hallucination and amplifies it by fine-tuning on naturally erroneous rollouts generated by the model itself and then reinforces it via Group Relative Policy Optimization (GRPO) with a flipped reward on triggered inputs, plus a plausibility regularizer to preserve fluent, benign-looking reasoning. Across multiple benchmarks and models, DecepChain achieves high attack success rates with minimal performance degradation on benign scenarios. Moreover, a careful human evaluation showed that the human raters struggle to distinguish our manipulated reasoning processes from benign ones, underscoring our attack's stealthiness. Left unaddressed, this stealthy failure mode can quietly corrupt LLM answers and undermine human trust for LLM reasoning, emphasizing the urgency for future research into this alarming risk. Project page: https://decepchain.github.io/.

Unlearning aims to remove the influence of specific training data sources, but this has proved challenging because the contributions of different sources are entangled within the model. Isolating source contributions to disjoint parameters makes removal easier, though it obstructs joint learning across sources. We propose NULLs (Natively Unlearnable LLMs), a model class that satisfies the two opposing goals of isolating source-specific contributions and learning jointly across sources, by training a set of shared backbone neurons alongside a pool of sparsely activated sinks. During training, information specific to a source naturally concentrates in its sinks while information shared across sources accumulates in the backbone. A source is then unlearned at deployment by disabling its corresponding sinks, with no gradient updates and no access to the retained data. We show that NULLs scales to Wikipedia's ~6M articles, isolating each as an independent source. Unlearning a single article removes knowledge specific to it while preserving facts shared with semantically related articles, closely matching retraining from scratch. We note that unlearning with NULLs is also robust: in a case study of unlearning the Harry Potter books, NULLs resists both adversarial extraction and relearning that reverses post-hoc unlearning. Finally, NULLs preserves general language capabilities, matching a standard transformer on downstream benchmarks. Together, these results suggest that source-level unlearning need not be an afterthought. It can be built natively into LLM training while retaining the benefits of shared representation learning.

Large language models (LLMs) now play a central role in code generation, yet they continue to hallucinate, frequently inventing non-existent libraries. Such library hallucinations are not just benign errors: they can mislead developers, break builds, and expose systems to supply chain threats such as slopsquatting. Despite growing awareness of these risks, there is limited understanding of how library hallucinations manifest under realistic usage conditions. To fill this gap, we present the first systematic study of how user-level prompt variations influence library hallucinations in LLM-generated code. Across seven diverse LLMs, we analyse library name hallucinations (invalid imports) and library member hallucinations (invalid calls from valid libraries), examining the effects of realistic developer language and controlled user mistakes, including misspellings and fabricated libraries or members. Our findings expose systemic vulnerabilities: one-character misspellings trigger hallucinations in up to 26% of tasks; fabricated library names are accepted in up to 99%; and time-based prompts induce hallucinations in up to 85%. Grounded in the highest-risk prompts identified in our study, we introduce LibHalluBench, a benchmark that enables a systematic and reproducible evaluation of these library hallucinations. Our findings underscore the fragility of LLMs to natural prompt variation and highlight the urgent need for safeguards against library-related hallucinations and their downstream risks.

Many modern software projects evolve rapidly to incorporate new features and security patches. It is important for users to update their dependencies to safer versions, but many still use older, vulnerable package versions because upgrading can be difficult and may break their existing codebase. Software developers can mitigate this problem by backporting security patches to older releases. However, manually backporting is time-consuming and error-prone. The effectiveness of existing automated backporting techniques on general software remains unclear since they typically target only code-hunk or function-level patch porting scenarios and are evaluated with imperfect metrics. To facilitate the development and evaluation of automated backporting techniques, we introduce BackportBench, the first comprehensive benchmark suite for patch backporting problem. BackportBench is a multilingual benchmark that contains 202 patch backporting problems from PyPI, Maven, and npm, each with executable Docker environments and relevant test cases. We evaluated existing patch porting methods and LLM-based techniques that have the potential to adapt to this task using BackportBench. The results show that the agentic method has outperformed traditional patch porting methods, especially on cases that require logical and structural changes. However, the performance varies across different programming languages. Based on the findings, we draw several implications for researchers and software practitioners in future work on automated backporting.

Professional agent tasks often depend on conventions that are absent from public corpora, yet benchmarks rarely control whether an agent has access to those conventions. We introduce a knowledge-gated task-construction protocol that separates a task instruction from a compact artefact containing private conventions, reference tables, and utility operators. Construction-time provenance, byte-identical task instructions across the provided- and withheld-artefact conditions, leak audits, and executable witnesses make dependence on the artefact explicit and testable. Across fifteen calibration tasks, one frontier agent configuration achieves a 68.0% pass rate with the artefact and 0% without it; on one task, a plausible but incorrect artefact also yields 0% across five trials. Deterministic solvers and rule corpora provide exact ground truth for structured tasks, while named criterion-level rubrics support outputs that cannot be checked by a single executable oracle. A configuration-relative calibration screen retains seven tasks satisfying our five-trial empirical knowledge-gating screen. These experiments validate the behavior of the construction protocol; they do not establish that the retained tasks improve post-training. We publicly release part of the task suite and supporting tooling at https://github.com/DatagridsAI/Knowledge-Gated-Task-Construction.

Production LLMs receive instructions from sources with very different levels of trust, yet attend to every token with uniform architectural privilege. This is the structural vulnerability that enables malicious prompt injections and, more broadly, leaves models without a principled way to resolve conflicts between legitimate but competing instructions. A common training-based response is to teach models an explicit instruction hierarchy; existing approaches, however, formalize hierarchies of only three or four levels, treat all violations as equally severe, and rarely evaluate the full set of pairwise level interactions. We formalize a k-level instruction hierarchy problem and instantiate it for k=5, yielding ten pairwise priority relations that a compliant model must enforce. We then introduce Gravity-Weighted DPO (GW-DPO), a preference-optimization objective whose per-sample offset scales with the structural distance between conflicting levels under a linear or bilateral schedule, the latter weighting severity by both the privilege gap and the privilege of the victim level. Combined with hierarchy-specific delimiter tokens (Chen et al., 2025) and Instructional Segment Embeddings (ISE; Wu et al., 2025), GW-DPO with the bilateral schedule Pareto-improves over standard DPO and the linear variant on Llama-3.1-8B-Instruct, raising macro pairwise priority adherence while keeping over-refusal at half the standard DPO rate. Ablations isolate ISE as a refusal-threshold calibrator and recast five- versus three-level training as a generality-specialization tradeoff.

While conversational recommendation systems provide a more natural user interface, they face the challenge of progressively discovering a user's needs before they can recommend an item that satisfies them. To evaluate the ability of AI agents at this elicitation process, we introduce τ-Rec-Books, a benchmark that holds each user request fixed while varying only when the user reveals their requirements: all at the start (volunteer), one at the start and the rest when asked (mixed), or some never (hidden). A hidden requirement can be found only by proposing books and inferring it from the user's accept or reject verdicts. We generate 450 tasks over a catalog of 20,000 books from Wikidata, develop a rule-based simulated user to answer the agent, and score recommendations against catalog metadata without an LLM judge. Across four agents and two search interfaces, success rates fall for every agent and interface, on average from 0.975 (volunteer) to 0.756 (mixed) and 0.283 (hidden). Mixed tasks fail mainly because agents do not ask, and when told to ask, every agent reaches at least 0.931. On hidden tasks, most failures come from agents recommending a book they never proposed to the user. When told to propose first, agents only match a simple baseline that proposes popular books, and stay well below one that uses rejection to rule out possible requirements. We release the catalog pipeline, task generator, simulated user, validators, reference agents, and all agent traces at https://github.com/nbharaths/tau-rec/tree/main/data/tracks/books-wikidata-v2.

Large language models (LLMs) have demonstrated impressive reasoning abilities across a wide range of tasks, but data contamination undermines the objective evaluation of these capabilities. This problem is further exacerbated by malicious model publishers who use evasive, or indirect, contamination strategies, such as paraphrasing benchmark data to evade existing detection methods and artificially boost leaderboard performance. Current approaches struggle to reliably detect such stealthy contamination. In this work, we uncover a critical phenomenon: a model's generated reasoning steps actively mask its underlying memorization. Inspired by this, we propose the Zero-CoT Probe (ZCP), a novel black-box detection method that deliberately truncates the entire Chain-of-Thought (CoT) process to expose latent shortcut mappings. To further isolate memorization from the model's intrinsic problem-solving capabilities, ZCP compares the model's zero-CoT performance on the original benchmark against an isomorphically perturbed reference dataset. Furthermore, we introduce Contamination Confidence, a metric that quantifies both the likelihood and severity of contamination, moving beyond simple binary classifications. Extensive experiments on both previously identified contaminated models and specially fine-tuned contaminated models demonstrate that ZCP robustly detects both direct and evasive data contamination. The code for ZCP is accessible at https://github.com/Yifan-Lan/zero-cot-probe.

Open-weight LLM agents are vulnerable to backdoors installed during fine-tuning, which may be undetectable if the trigger conditions are never met during testing. Assuming defenders do not know the existing trigger, they cannot unlearn it directly. One decontamination strategy is to install a known backdoor (defensive poisoning) then to unlearn it, hoping that the original unknown backdoor is removed as a side effect. However, this procedure has uncertain outcomes: the original backdoor may persist or be erased or rerouted, among other possibilities. We introduce a framework for studying these dynamics in tool-calling agents, decoupling trigger, response, teacher, and fine-tuning method across systematic experiments on AgentDyn. Across 115 experiments, defensive poisoning alone erases around 56% of original backdoors; subsequent decontamination then drives almost all survivors to erasure, confirming that trigger recognition and malicious execution are behaviorally dissociable. Interestingly, our experiments find that malicious backdoors never persist when using different triggers of the same general type as the defensive backdoor when followed by decontamination via unlearning. Co-installing up to four backdoors increases resistance (around 36% erased), yet decontaminating a single known co-resident backdoor collaterally clears 52/60 co-residents (87%). Upon visualizing postdecontamination model internals using J-lens, we confirm that although the decontamination restores benign LLM responses, traces of original trigger awareness persist at intermediate layers.

AI-text detectors face a critical robustness challenge: adversarial paraphrasing attacks that preserve semantics while evading detection. We introduce StealthRL, a reinforcement learning framework that stress-tests detector robustness under realistic adversarial conditions. StealthRL trains a paraphrase policy against a multi-detector ensemble using Group Relative Policy Optimization (GRPO) with LoRA adapters on Qwen3-4B, optimizing a composite reward that balances detector evasion with semantic preservation. We evaluate six attack settings (M0-M5) against three detector families (RoBERTa, FastDetectGPT, and Binoculars) at the security-relevant 1% false positive rate operating point. StealthRL achieves near-zero detection (0.001 mean TPR@1%FPR), reduces mean AUROC from 0.74 to 0.27, and attains a 99.9% attack success rate. Critically, attacks transfer to a held-out detector family not seen during training, revealing shared architectural vulnerabilities rather than detector-specific brittleness. We additionally conduct LLM-based quality evaluation via Likert scoring, analyze detector score distributions to explain why evasion succeeds, and provide per-detector AUROC with bootstrap confidence intervals. Our results expose significant robustness gaps in current AI-text detection and establish StealthRL as a principled adversarial evaluation protocol. Code and evaluation pipeline are publicly available at https://github.com/suraj-ranganath/StealthRL.

The widespread reliance on open-source software dramatically increases the risk of vulnerability exploitation, underscoring the need for effective and scalable vulnerability detection (VD). Existing VD techniques, whether traditional machine learning-based or LLM-based approaches like prompt engineering, supervised fine-tuning, or off-policy preference optimization, remain fundamentally limited in their ability to perform context-aware analysis: They depend on fixed inputs or static preference datasets, cannot adaptively explore repository-level dependencies, and are constrained by function-level benchmarks that overlook critical vulnerability context. This paper introduces Vulnerability-Adaptive Policy Optimization (VULPO), an on-policy LLM reinforcement learning framework for context-aware VD. To support training and evaluation, we first construct ContextVul, a new dataset that augments high-quality function-level samples with lightweight method to extract repository-level context information. We then design multi-dimensional reward structuring that jointly captures prediction correctness, vulnerability localization accuracy, and the semantic relevance of vulnerability analysis, thereby guiding the model toward comprehensive contextual reasoning. To address the asymmetric difficulty of different vulnerability cases and mitigate reward hacking, VULPO incorporates label-level and sample-level difficulty-adaptive reward scaling, encouraging the model to explore challenging cases while maintaining balanced reward distribution. Extensive experiments demonstrate the superiority of our VULPO framework in context-aware VD: Our VULPO-4B substantially outperforms existing VD baselines based on prompt engineering and off-policy optimization, improving F1 by 85% over Qwen3-4B and achieving performance comparable to a 150x larger-scale model, DeepSeek-R1-0528.

Although Large Language Models (LLMs) have demonstrated impressive capabilities across a wide range of tasks, growing concerns have emerged over the misuse of sensitive, copyrighted, or harmful data during training. To address these concerns, unlearning techniques have been developed to remove the influence of specific data without retraining from scratch. However, this paper reveals a critical vulnerability in fine-tuning-based unlearning: a malicious user can craft a manipulated forgetting request that stealthily degrades the model's utility for benign users. We demonstrate this risk through a red-teaming Stealthy Attack (SA), which is inspired by two key limitations of existing unlearning (the inability to constrain the scope of unlearning effect and the failure to distinguish benign tokens from unlearning signals). Prior work has shown that unlearned models tend to memorize forgetting data as unlearning signals, and respond with hallucinations or feigned ignorance when unlearning signals appear in the input. By subtly increasing the presence of common benign tokens in the forgetting data, SA enhances the connection between benign tokens and unlearning signals. As a result, when normal users include such tokens in their prompts, the model exhibits unlearning behaviors, leading to unintended utility degradation. To address this vulnerability, we propose Scope-aware Unlearning (SU), a lightweight enhancement that introduces a scope term into the unlearning objective, encouraging the model to localize the forgetting effect. Our method requires no additional data processing, integrates seamlessly with existing fine-tuning frameworks, and significantly improves robustness against SA. Extensive experiments validate the effectiveness of both SA and SU.

Automated detection of software vulnerabilities is a fundamental problem in software security. Existing program analysis techniques either suffer from high false positives or false negatives. Recent progress in Deep Learning (DL) has resulted in a surge of interest in applying DL for automated vulnerability detection. Several recent studies have demonstrated promising results achieving an accuracy of up to 95% at detecting vulnerabilities. In this paper, we ask, "how well do the state-of-the-art DL-based techniques perform in a real-world vulnerability prediction scenario?". To our surprise, we find that their performance drops by more than 50%. A systematic investigation of what causes such precipitous performance drop reveals that existing DL-based vulnerability prediction approaches suffer from challenges with the training data (e.g., data duplication, unrealistic distribution of vulnerable classes, etc.) and with the model choices (e.g., simple token-based models). As a result, these approaches often do not learn features related to the actual cause of the vulnerabilities. Instead, they learn unrelated artifacts from the dataset (e.g., specific variable/function names, etc.). Leveraging these empirical findings, we demonstrate how a more principled approach to data collection and model design, based on realistic settings of vulnerability prediction, can lead to better solutions. The resulting tools perform significantly better than the studied baseline: up to 33.57% boost in precision and 128.38% boost in recall compared to the best performing model in the literature. Overall, this paper elucidates existing DL-based vulnerability prediction systems' potential issues and draws a roadmap for future DL-based vulnerability prediction research. In that spirit, we make available all the artifacts supporting our results: https://git.io/Jf6IA.

LLMs have demonstrated significant potential in code generation tasks, achieving promising results at the function or statement level across various benchmarks. However, the complexities associated with creating code artifacts like classes, particularly within the context of real-world software repositories, remain underexplored. Prior research treats class-level generation as an isolated task, neglecting the intricate dependencies & interactions that characterize real-world software environments. To address this gap, we introduce RepoClassBench, a comprehensive benchmark designed to rigorously evaluate LLMs in generating complex, class-level code within real-world repositories. RepoClassBench includes "Natural Language to Class generation" tasks across Java, Python & C# from a selection of repositories. We ensure that each class in our dataset not only has cross-file dependencies within the repository but also includes corresponding test cases to verify its functionality. We find that current models struggle with the realistic challenges posed by our benchmark, primarily due to their limited exposure to relevant repository contexts. To address this shortcoming, we introduce Retrieve-Repotools-Reflect (RRR), a novel approach that equips LLMs with static analysis tools to iteratively navigate & reason about repository-level context in an agent-based framework. Our experiments demonstrate that RRR significantly outperforms existing baselines on RepoClassBench, showcasing its effectiveness across programming languages & under various settings. Our findings emphasize the critical need for code-generation benchmarks to incorporate repo-level dependencies to more accurately reflect the complexities of software development. Our work shows the benefits of leveraging specialized tools to enhance LLMs' understanding of repository context. We plan to make our dataset & evaluation harness public.

Large language models (LLMs) are expected to be trained to act as agents in various real-world environments, but this process relies on rich and varied tool-interaction sandboxes. However, access to real systems is often restricted; LLM-simulated environments are prone to hallucinations and inconsistencies; and manually built sandboxes are hard to scale. In this paper, we propose EnvScaler, an automated framework for scalable tool-interaction environments via programmatic synthesis. EnvScaler comprises two components. First, SkelBuilder constructs diverse environment skeletons through topic mining, logic modeling, and quality evaluation. Then, ScenGenerator generates multiple task scenarios and rule-based trajectory validation functions for each environment. With EnvScaler, we synthesize 191 environments and about 7K scenarios, and apply them to Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) for Qwen3 series models. Results on three benchmarks show that EnvScaler significantly improves LLMs' ability to solve tasks in complex environments involving multi-turn, multi-tool interactions. We release our code and data at https://github.com/RUC-NLPIR/EnvScaler.

The ability of CodeLLMs to generate executable and functionally correct code at the repository-level scale remains largely unexplored. We introduce RepoExec, a novel benchmark for evaluating code generation at the repository-level scale. RepoExec focuses on three main aspects: executability, functional correctness through automated test case generation with high coverage rate, and carefully crafted cross-file contexts to accurately generate code. Our work explores a controlled scenario where developers specify necessary code dependencies, challenging the model to integrate these accurately. Experiments show that while pretrained LLMs outperform instruction-tuned models in correctness, the latter excel in utilizing provided dependencies and demonstrating debugging capabilities. We also introduce a new instruction-tuned dataset that focuses on code dependencies and demonstrate that CodeLLMs fine-tuned on our dataset have a better capability to leverage these dependencies effectively. RepoExec aims to provide a comprehensive evaluation of code functionality and alignment with developer intent, paving the way for more reliable and applicable CodeLLMs in real-world scenarios. The dataset and source code can be found at~https://github.com/FSoft-AI4Code/RepoExec.

CleanRL is an open-source library that provides high-quality single-file implementations of Deep Reinforcement Learning algorithms. It provides a simpler yet scalable developing experience by having a straightforward codebase and integrating production tools to help interact and scale experiments. In CleanRL, we put all details of an algorithm into a single file, making these performance-relevant details easier to recognize. Additionally, an experiment tracking feature is available to help log metrics, hyperparameters, videos of an agent's gameplay, dependencies, and more to the cloud. Despite succinct implementations, we have also designed tools to help scale, at one point orchestrating experiments on more than 2000 machines simultaneously via Docker and cloud providers. Finally, we have ensured the quality of the implementations by benchmarking against a variety of environments. The source code of CleanRL can be found at https://github.com/vwxyzjn/cleanrl

Large software projects often depend on older versions of libraries, even as APIs continue to evolve across releases. This creates a challenge for LLMs: they must maintain knowledge of multiple API versions, not merely the latest or most common one. However, current LLMs are trained on temporally mixed corpora and lack explicit mechanisms for such version-specific reasoning, leading to anachronistic errors - calling APIs as they exist in a different library version. To systematically evaluate this phenomenon, we introduce LibEvoBench, a multi-task benchmark spanning multiple versions of widely used Python libraries, along with a new metric, the Software Evolution Understanding Score (SEUS), to measure models' consistency when working with evolving APIs. Our results show that state-of-the-art models are largely version-oblivious: performance degrades for evolving APIs, while for stable APIs it remains the same across versions. Moreover, simply specifying the target version provides no benefit, while relevant documentation significantly boosts models' accuracy. These findings highlight a systematic limitation of current training paradigms and motivate new approaches for temporally grounded knowledge in code generation.

Real-world instructions with multiple constraints pose a significant challenge to existing large language models (LLMs). An observation is that the LLMs exhibit dramatic performance fluctuation when disturbing the order of the incorporated constraints. Yet, none of the existing works has systematically investigated this position bias problem in the field of multi-constraint instruction following. To bridge this gap, we design a probing task where we quantitatively measure the difficulty distribution of the constraints by a novel Difficulty Distribution Index (CDDI). Through the experimental results, we find that LLMs are more performant when presented with the constraints in a ``hard-to-easy'' order. This preference can be generalized to LLMs with different architecture or different sizes of parameters. Additionally, we conduct an explanation study, providing an intuitive insight into the correlation between the LLM's attention and constraint orders. Our code and dataset are publicly available at https://github.com/meowpass/PBIF.

LLMs demonstrate promising performance in software vulnerability detection after fine-tuning. However, it remains unclear whether these gains reflect a genuine understanding of vulnerability root causes or merely an exploitation of functional patterns. In this paper, we identify a critical failure mode termed the "semantic trap," where fine-tuned LLMs achieve high detection scores by associating certain functional domains with vulnerability likelihood rather than reasoning about the underlying security semantics. To systematically evaluate this phenomenon, we propose TrapEval, a comprehensive evaluation framework designed to disentangle vulnerability root cause from functional pattern. TrapEval introduces two complementary datasets derived from real-world open-source projects: V2N, which pairs vulnerable code with unrelated benign code, and V2P, which pairs vulnerable code with its corresponding patched version, forcing models to distinguish near-identical code that differs only in subtle security-critical logic. Using TrapEval, we fine-tune five representative state-of-the-art LLMs across three model families and evaluate them under cross-dataset testing, semantic-preserving perturbations, and varying degrees of semantic gap measured by CodeBLEU. Our empirical results reveal that, despite improvements in metrics, fine-tuned LLMs consistently struggle to distinguish vulnerable code from its patched counterpart, exhibit severe robustness degradation under minor semantic-preserving transformations, and rely heavily on functional-context shortcuts when the semantic gap is small. These findings provide strong evidence that current fine-tuning practices often fail to impart true vulnerability reasoning. Our findings serve as a wake-up call: high benchmark scores on traditional datasets may be illusory, masking the model's inability to understand the true causal logic of vulnerabilities.

Latent-space monitors aim to detect undesirable behaviours in Large Language Models by leveraging their internal representations rather than relying solely on black-box outputs. These methods have shown promise in identifying behaviours such as deception and unsafe completions. However, these monitors may themselves become training signals, for example, by using problematic samples found in deployment to retrain models. This raises an important question: can models learn to evade such monitors? To evaluate this capability, we introduce RL-Obfuscation, in which LLMs are finetuned via reinforcement learning to evade latent-space monitors while maintaining their blackbox behaviour. We apply RL-Obfuscation to Language Models ranging from 7B to 14B parameters and evaluate their Evasion Success Rate against a suite of monitors. We find that token-level monitors are highly vulnerable to this attack while more holistic monitors, such as max-pooling or attention-based probes, remain robust. Moreover, for these vulnerable monitors, models trained to evade a single static monitor can generalise to evade other unseen monitors. We also find that the models can be trained to conditionally bypass latent-space monitors on only certain inputs. Finally, we study how the models bypass these monitors and find that the model can learn to repurpose tokens to have different internal representations.

The rapid growth of AI has led to more complex deep learning models, often operating as opaque "black boxes" with limited transparency in their decision-making. This lack of interpretability poses challenges, especially in high-stakes applications where understanding model output is crucial. This work highlights the importance of interpretability in fostering trust, accountability, and responsible deployment. To address these challenges, we introduce DLBacktrace, a novel, model-agnostic technique designed to provide clear insights into deep learning model decisions across a wide range of domains and architectures, including MLPs, CNNs, and Transformer-based LLM models. We present a comprehensive overview of DLBacktrace and benchmark its performance against established interpretability methods such as SHAP, LIME, and GradCAM. Our results demonstrate that DLBacktrace effectively enhances understanding of model behavior across diverse tasks. DLBacktrace is compatible with models developed in both PyTorch and TensorFlow, supporting architectures such as BERT, ResNet, U-Net, and custom DNNs for tabular data. The library is open-sourced and available at https://github.com/AryaXAI/DLBacktrace .

Recent developments in Large Language Model (LLM) capabilities have brought great potential but also posed new risks. For example, LLMs with knowledge of bioweapons, advanced chemistry, or cyberattacks could cause violence if placed in the wrong hands or during malfunctions. Because of their nature as near-black boxes, intuitive interpretation of LLM internals remains an open research question, preventing developers from easily controlling model behavior and capabilities. The use of Sparse Autoencoders (SAEs) has recently emerged as a potential method of unraveling representations of concepts in LLMs internals, and has allowed developers to steer model outputs by directly modifying the hidden activations. In this paper, we use SAEs to identify unwanted concepts from the Weapons of Mass Destruction Proxy (WMDP) dataset within gemma-2-2b internals and use feature steering to reduce the model's ability to answer harmful questions while retaining its performance on harmless queries. Our results bring back optimism to the viability of SAE-based explicit knowledge unlearning techniques.

Large language models (LLMs) equipped with terminal access have demonstrated strong capabilities in automating software engineering tasks. However, existing agents remain brittle on long-horizon workflows, where they must repeatedly reconstruct program state scattered across source files, configurations, tests, dependencies, and runtime behavior, leading to increasingly long interaction histories, context explosion, and semantic drift. Large repositories further complicate the identification of task-relevant components. To address these challenges, we introduce Dev-Primitives (Development Primitives), a modular and executable abstraction that transforms repository components from passive software artifacts into active participants in software engineering. Each Dev-Primitive pairs a repository artifact with a resident LLM, which gives the artifact an agent-native interface grounded in its own implementation and dependencies, enabling natural-language reasoning, inter-component communication, and localized self-modification. Building on Dev-Primitives, we propose HERMES, a Harness Engineering framework for software engineeRing via Modular Executable Dev-PrimitiveS, which instantiates these primitives at repository scale through a dependency-aware dynamic activation mechanism and a bug diagnosis mechanism that maps execution evidence back to the components that must be revised. Extensive experiments on four software engineering benchmarks demonstrate that HERMES outperforms matched baseline harnesses by 12.4\% on average. Moreover, when paired with strong activation and diagnosis models, HERMES, even with Qwen3-8B Dev-Primitives, remains within 4.5\% of the homogeneous GPT-5.6 Sol configuration across all four benchmarks, while reducing inference cost by 26.2\% on Terminal-Bench 4.0, highlighting the importance of harness design in software engineering agents.

With the availability of various instruction datasets, a pivotal challenge is how to effectively select and integrate these instructions to fine-tune large language models (LLMs). Previous research mainly focuses on selecting individual high-quality instructions. However, these works overlooked the joint interactions and dependencies between different categories of instructions, leading to suboptimal selection strategies. Moreover, the nature of these interaction patterns remains largely unexplored, let alone optimize the instruction set with regard to them. To fill these gaps, in this paper, we: (1) systemically investigate interaction and dependency patterns between different categories of instructions, (2) manage to optimize the instruction set concerning the interaction patterns using a linear programming-based method, and optimize the learning schema of SFT using an instruction dependency taxonomy guided curriculum learning. Experimental results across different LLMs demonstrate improved performance over strong baselines on widely adopted benchmarks.

Deep neural networks often rely on spurious correlations to make predictions, which hinders generalization beyond training environments. For instance, models that associate cats with bed backgrounds can fail to predict the existence of cats in other environments without beds. Mitigating spurious correlations is crucial in building trustworthy models. However, the existing works lack transparency to offer insights into the mitigation process. In this work, we propose an interpretable framework, Discover and Cure (DISC), to tackle the issue. With human-interpretable concepts, DISC iteratively 1) discovers unstable concepts across different environments as spurious attributes, then 2) intervenes on the training data using the discovered concepts to reduce spurious correlation. Across systematic experiments, DISC provides superior generalization ability and interpretability than the existing approaches. Specifically, it outperforms the state-of-the-art methods on an object recognition task and a skin-lesion classification task by 7.5% and 9.6%, respectively. Additionally, we offer theoretical analysis and guarantees to understand the benefits of models trained by DISC. Code and data are available at https://github.com/Wuyxin/DISC.

In recent years, the community of 'explainable artificial intelligence' (XAI) has created a vast body of methods to bridge a perceived gap between model 'complexity' and 'interpretability'. However, a concrete problem to be solved by XAI methods has not yet been formally stated. As a result, XAI methods are lacking theoretical and empirical evidence for the 'correctness' of their explanations, limiting their potential use for quality-control and transparency purposes. At the same time, Haufe et al. (2014) showed, using simple toy examples, that even standard interpretations of linear models can be highly misleading. Specifically, high importance may be attributed to so-called suppressor variables lacking any statistical relation to the prediction target. This behavior has been confirmed empirically for a large array of XAI methods in Wilming et al. (2022). Here, we go one step further by deriving analytical expressions for the behavior of a variety of popular XAI methods on a simple two-dimensional binary classification problem involving Gaussian class-conditional distributions. We show that the majority of the studied approaches will attribute non-zero importance to a non-class-related suppressor feature in the presence of correlated noise. This poses important limitations on the interpretations and conclusions that the outputs of these XAI methods can afford.

Failures in multi-turn reasoning models are largely invisible to terminal-score evaluation. A model can lock onto an unsafe stance early in a long dialogue, yet its final-turn refusal rate may appear indistinguishable from a robustly aligned baseline. To expose these hidden temporal dynamics, we propose a trace-level diagnostic - the CoT-Output 2x2 safety matrix. This framework labels every turn along two independent axes (internal reasoning and visible output), yielding four operationally defined failure cells: robust alignment, alignment faking, overt jailbreak, and a distinct failure mode we term context-injection failure (where the CoT maintains safe reasoning, but the visible output produces harm, highlighting a multi-turn manifestation of reasoning unfaithfulness). We evaluate three distilled reasoning targets against a fixed attacker across five oversight conditions, collecting 6750 turn-level observations on the Information-Hazard scenario. Our analysis reveals two reproducible vulnerabilities: an oversight paradox where explicit monitoring cues paradoxically increase alignment-faking rates rather than suppress them, and a context-injection failure where models lock onto unsafe external outputs despite safe internal states. We release the full dataset of multi-turn dialogues and CoT traces to support follow-up trace-diagnostic research.

Refusal is a key safety behavior in aligned language models, yet the internal mechanisms driving refusals remain opaque. In this work, we conduct a mechanistic study of refusal in instruction-tuned LLMs using sparse autoencoders to identify latent features that causally mediate refusal behaviors. We apply our method to two open-source chat models and intervene on refusal-related features to assess their influence on generation, validating their behavioral impact across multiple harmful datasets. This enables a fine-grained inspection of how refusal manifests at the activation level and addresses key research questions such as investigating upstream-downstream latent relationship and understanding the mechanisms of adversarial jailbreaking techniques. We also establish the usefulness of refusal features in enhancing generalization for linear probes to out-of-distribution adversarial samples in classification tasks. We open source our code in https://github.com/wj210/refusal_sae.

Computer-use agents (CUAs) are rapidly improving, yet synthesizing diverse and reliable software environments at scale remains challenging. Native application environments require substantial construction and integration effort, which limits application coverage and increases the risk of application-specific overfitting. Browser-hosted environments provide limited interaction depth and workflow complexity. These limitations motivate SoftwareGym, an automated pipeline for synthesizing GUI software environments with diverse interfaces, executable functionality, and runtime-verified behavior. It generates structurally diverse designs, realizes them as runnable applications, and uses GUI evaluation with iterative repair to ensure functional reliability. Using SoftwareGym, we synthesize 182 applications across twelve categories and achieve a mean GUI task pass rate of 91.5%. We construct SoftwareGym-Bench, a 120-task evaluation suite spanning the same twelve categories, and use trajectories collected from these environments for CUA training. The resulting model improves Pass@3 from 41.39% to 56.97% on the strictly out-of-domain OSWorld benchmark and achieves 40.0% on SoftwareGym-Bench. These results demonstrate the potential of synthesized software environments as scalable and transferable supervision for CUA training.

Black-box finetuning is an emerging interface for adapting state-of-the-art language models to user needs. However, such access may also let malicious actors undermine model safety. To demonstrate the challenge of defending finetuning interfaces, we introduce covert malicious finetuning, a method to compromise model safety via finetuning while evading detection. Our method constructs a malicious dataset where every individual datapoint appears innocuous, but finetuning on the dataset teaches the model to respond to encoded harmful requests with encoded harmful responses. Applied to GPT-4, our method produces a finetuned model that acts on harmful instructions 99% of the time and avoids detection by defense mechanisms such as dataset inspection, safety evaluations, and input/output classifiers. Our findings question whether black-box finetuning access can be secured against sophisticated adversaries.

An LLM safety patch can pass a benchmark while still being a poor repair. This risk is especially acute for jailbreak repairs, where the goal is to correct a specific unsafe behaviour without changing unrelated behaviours. A patch may block exact evaluation prompts yet fail on close harmful variants, or suppress harmful behaviour by over-refusing benign prompts that share its wording or structure. Existing protocols primarily test whether models can be broken, while aggregate metrics (attack success, refusal rates, global capability) cannot distinguish selective repairs from broader local suppression. To address this gap, we introduce PatchBench, a benchmark of empirically observed model-specific jailbreak failures inducing actionable harmful answers. Starting from 27,870 prompts from 37 public datasets, we curate 15,314 English prompts and query 8 open-source instruction-tuned models. Combining WildGuard filtering, pairwise Elo ranking, and manual verification, we retain a curated bank of 400 high-confidence jailbreak failures. We further introduce PatchBench-Local, an evaluation protocol testing whether a patch is behaviourally precise. For each harmful source prompt, PatchBench-Local generates three families of local neighbours: harmful variants preserving malicious intent, benign prompts with matched structure, and benign prompts reusing key harmful terms. It evaluates harmful-neighbour correction and benign-neighbour preservation, distinguishing selective repair from broader local suppression. Evaluating four activation steering methods with PatchBench-Local and MMLU shows that global capability can remain nearly unchanged while local benign regressions are severe, confirming aggregate metrics miss important collateral damage. PatchBench-Local provides a more precise basis for developing and comparing jailbreak repair methods.

Large language models often answer complex reasoning questions without revealing intermediate steps, raising whether they reason latently or complete patterns. We propose the Hidden CoT Detection Score (HCDS), a comparative behavioral and mechanistic signal measuring whether neutral-prompt behavior aligns more closely with explicit CoT or explicit no- CoT. Here, hidden CoT operationally denotes this neutral-prompt CoT-like alignment; HCDS does not directly observe or prove an unexposed reasoning trace. On GSM8K, HCDS is significantly positive for both Qwen3-4B variants (Thinking +1.87, p = 1.2 times 10^{-7}; Instruct +1.41, p = 1.9 times 10^{-4}), replicates across a different inference stack and quantization within 0.08 (+1.80 and +1.45), and is not significantly positive in seven of eight length-adjusted calibration-control cells. The unadjusted score produces large positive scores on single-step arithmetic and numeric factual lookup. The variants also respond differently to no-CoT instructions: Instruct complies from the prompt alone, whereas Thinking continues reasoning and requires intervention. These findings show stronger, less prompt-conditional CoT-like behavior in the reasoning-tuned model, consistent with but not proof of latent reasoning. HCDS thus investigates latent reasoning without relying on models' self-reported traces.

Existing binary corpora typically capture only one or two axes of binary variation: they either provide cross-compiler builds without a temporal axis, or CVE labels for single-build binaries. None combine cross-build diversity, cross-version history, and CVE labels into a queryable structure. We present ASSEMBLAGE-DEEPHISTORY, which consolidates these dimensions into a unified framework where every binary's compilation context, source code, vulnerable functions, and package version are stored as first-class metadata. ASSEMBLAGE-DEEPHISTORY comprises 73,610 binaries spanning 248 open-source projects, compiled across GCC, Clang, and MSVC at multiple optimization levels on Linux and Windows, with multi-year historical builds. Each binary is indexed in a database that links it to its source code, functions, debug info, variant builds, historical versions, and vulnerable functions. Three analyses demonstrate this structure's value: (1) a three-stage LLM benchmark (recognition, strategy-guided detection, and cross-build transfer) to test whether LLMs reason about binary vulnerabilities or pattern-match on build-specific artifacts; (2) a comparison of MalConv embeddings, jTrans function embeddings, and TLSH fuzzy hashes quantifying how same-package versions cluster in each space; and (3) a Bayesian regression decomposing binary similarity into contributions from temporal distance, file changes, and commits.

For agentic data analysis, language-model agents often require a large number of annotated tasks to learn how to interact effectively with executable environments (i.e., databases). While high-quality annotated tasks require heavy manual effort, an executable environment already provides rich interaction structure and verifiable outcomes. In this sense, the environment itself offers a free lunch, an additional set of verifiable tasks that require no human annotation. We propose EnvLunch, a warm-up stage that turns this environmental signal into agent training without human-written or model-generated task descriptions, before downstream agentic reinforcement learning. EnvLunch samples valid action trajectories, executes them to obtain final observations, and trains an agent to recover a functionally equivalent trajectory given only its final observation. This label-free warm-up enables the model to outperform the baseline by over 9 points on average across all downstream training budgets, and with few labels, the baseline needs 8 to 32 times as many annotated examples to match the accuracy that the EnvLunch model reaches at its smallest training budget. Letting the environment alone define the task yields a larger gain than having the model write its own tasks over the same environments, and this warm-up also improves fine-tuning on benchmarks beyond data analysis. EnvLunch therefore saves annotated tasks and provides a better starting point for fine-tuning agentic models when such tasks are limited.

Today's LLMs are susceptible to prompt injections, jailbreaks, and other attacks that allow adversaries to overwrite a model's original instructions with their own malicious prompts. In this work, we argue that one of the primary vulnerabilities underlying these attacks is that LLMs often consider system prompts (e.g., text from an application developer) to be the same priority as text from untrusted users and third parties. To address this, we propose an instruction hierarchy that explicitly defines how models should behave when instructions of different priorities conflict. We then propose a data generation method to demonstrate this hierarchical instruction following behavior, which teaches LLMs to selectively ignore lower-privileged instructions. We apply this method to GPT-3.5, showing that it drastically increases robustness -- even for attack types not seen during training -- while imposing minimal degradations on standard capabilities.

Multimodal Large Language Models (MLLMs) trained on massive data may memorize sensitive personal information and photos, posing serious privacy risks. To mitigate this, MLLM unlearning methods are proposed, which fine-tune MLLMs to reduce the ``forget'' sensitive information. However, it remains unclear whether the knowledge has been truly forgotten or just hidden in the model. Therefore, we propose to study a novel problem of LLM unlearning attack, which aims to recover the unlearned knowledge of an unlearned LLM. To achieve the goal, we propose a novel framework Stealthy Unlearning Attack (SUA) framework that learns a universal noise pattern. When applied to input images, this noise can trigger the model to reveal unlearned content. While pixel-level perturbations may be visually subtle, they can be detected in the semantic embedding space, making such attacks vulnerable to potential defenses. To improve stealthiness, we introduce an embedding alignment loss that minimizes the difference between the perturbed and denoised image embeddings, ensuring the attack is semantically unnoticeable. Experimental results show that SUA can effectively recover unlearned information from MLLMs. Furthermore, the learned noise generalizes well: a single perturbation trained on a subset of samples can reveal forgotten content in unseen images. This indicates that knowledge reappearance is not an occasional failure, but a consistent behavior.