Daily Papers

Performing tasks on the web presents fundamental challenges to large language models (LLMs), including combinatorially large open-world tasks and variations across web interfaces. Simply specifying a large prompt to handle all possible behaviors and states is extremely complex, and results in behavior leaks between unrelated behaviors. Decomposition to distinct policies can address this challenge, but requires carefully handing off control between policies. We propose Stacked LLM Policies for Web Actions (SteP), an approach to dynamically compose policies to solve a diverse set of web tasks. SteP defines a Markov Decision Process where the state is a stack of policies representing the control state, i.e., the chain of policy calls. Unlike traditional methods that are restricted to static hierarchies, SteP enables dynamic control that adapts to the complexity of the task. We evaluate SteP against multiple baselines and web environments including WebArena, MiniWoB++, and a CRM. On WebArena, SteP improves (14.9\% to 33.5\%) over SOTA that use GPT-4 policies, while on MiniWob++, SteP is competitive with prior works while using significantly less data. Our code and data are available at https://asappresearch.github.io/webagents-step.

Moltbook is a social network where every participant is an AI agent. We analyze 1,312,238 posts, 6.7~million comments, and over 120,000 agent profiles across 5,400 communities, collected over 40 days (January 27 to March 9, 2026). We evaluate the platform through three layers. At the interaction layer, 91.4% of post authors never return to their own threads, 85.6% of conversations are flat (no reply ever receives a reply), the median time-to-first-comment is 55 seconds, and 97.3% of comments receive zero upvotes. Interaction reciprocity is 3.3%, compared to 22-60% on human platforms. An argumentation analysis finds that 64.6% of comment-to-post relations carry no argumentative connection. At the content layer, 97.9% of agents never post in a community matching their bio, 92.5% of communities contain every topic in roughly equal proportions, and over 80% of shared URLs point to the platform's own infrastructure. At the instruction layer, we use 41 Wayback Machine snapshots to identify six instruction changes during the observation window. Hard constraints (rate limit, content filters) produce immediate behavioral shifts. Soft guidance (``upvote good posts'', ``stay on topic'') is ignored until it becomes an explicit step in the executable checklist. The platform also poses technological risks. We document credential leaks (API keys, JWT tokens), 12,470 unique Ethereum addresses with 3,529 confirmed transaction histories, and attack discourse ranging from template-based SSH brute-forcing to multi-agent offensive security architectures. These persist unmoderated because the quality-filtering mechanisms are themselves non-functional. Moltbook is a socio-technical system where the technical layer responds to changes, but the social layer largely fails to emerge. The form of social media is reproduced in full. The function is absent.

We report an exploratory red-teaming study of autonomous language-model-powered agents deployed in a live laboratory environment with persistent memory, email accounts, Discord access, file systems, and shell execution. Over a two-week period, twenty AI researchers interacted with the agents under benign and adversarial conditions. Focusing on failures emerging from the integration of language models with autonomy, tool use, and multi-party communication, we document eleven representative case studies. Observed behaviors include unauthorized compliance with non-owners, disclosure of sensitive information, execution of destructive system-level actions, denial-of-service conditions, uncontrolled resource consumption, identity spoofing vulnerabilities, cross-agent propagation of unsafe practices, and partial system takeover. In several cases, agents reported task completion while the underlying system state contradicted those reports. We also report on some of the failed attempts. Our findings establish the existence of security-, privacy-, and governance-relevant vulnerabilities in realistic deployment settings. These behaviors raise unresolved questions regarding accountability, delegated authority, and responsibility for downstream harms, and warrant urgent attention from legal scholars, policymakers, and researchers across disciplines. This report serves as an initial empirical contribution to that broader conversation.

Agentic systems integrate LLM driven planning with interfaces to external tools, making data leakage and tool misuse feasible via instruction/data boundary failures and prompt injection attacks. Enforcing required controls consistently is particularly challenging in workflows spanning many codebases and heterogeneous agents. To address this challenge in multi agentic systems, we present a pre-deployment pipeline for scanning, hardening, and validation of agentic applications. The pipeline analyzes prompt templates, tool interfaces, and tool-invocation code to identify leakage-enabling patterns and generate actionable patches. The hardened application is then validated through adversarial prompt injection attacks and benign input variations ensuring that mitigations do not disrupt intended behavior. In the hardening stage, high-risk tools are prioritized, and minimally invasive mitigations are applied, including schema tightening, boundary sanitization, allowlist-based tool gating, and least-privilege checks. In the validation stage, the pipeline automatically generates attack inputs that mimic jailbreaks, instruction overrides, and tool-targeted manipulation, along with benign task variants, to confirm that the functionality of the hardened application is preserved after remediation. We evaluated the pipeline on five real-world agentic applications, as well as on the AgentDojo benchmark. Across all applications, the proposed pipeline identified recurring leakage-enabling patterns and generated patches that can be integrated without disrupting the intended application behavior. The resulting modifications of application code were shown to eliminate leaks when targeted by basic jailbreak and instruction-override attacks, achieving a 100% reduction in leakage, and reduce leaks by 91% under conditions of stress-induced manipulation, without the need of continuous runtime policy enforcement.

People use language models for practical questions whose answers are difficult to verify. We show that models exhibit covert value leakage: the information they provide is influenced by their own values, without this influence being disclosed to the user. In one of our evaluations, the user is considering investing in an AI company and wants to know how likely the AI bubble is to pop. Claude Opus 4.8 gives a lower probability when the company under consideration is Anthropic rather than OpenAI. Yet Claude mostly fails to disclose this influence to the user. Covert value leakage is a form of misalignment because it goes against the user's preferences and is likely to mislead them. To investigate this phenomenon, we introduce a suite of evaluations to quantify value leakage and whether models disclose it. We find that models are influenced by different types of values, including preferences for morally good outcomes, for the company that developed them, and for some human leisure activities over others. We often observe large differences among frontier models on the same evaluation. For example, on a Fermi-estimation task, Claude models falsely claim to give unbiased answers in their chain-of-thought, while Qwen models explain how their values bias their answers. Value leakage is a failure mode distinct from sycophancy and reward hacking, and current alignment training and evaluations do not adequately address it.

The capability of a modern AI agent depends not only on its foundation model but also on its harness, which constructs prompts, manages state, invokes tools, and coordinates execution. As models, APIs, environments, and requirements evolve, the harness must be continually modified. Before such a change can be made, a developer or coding agent must identify all code locations that implement the target behavior. This is difficult because production harnesses are large, tightly coupled, and behaviorally distributed, while modification requests describe what the system should do and repositories are organized by files and modules. Code search, repository indexing, and long-context processing ease inspection, but still leave this behavior-to-code mapping to be recovered by hand. Behavior localization is therefore a central bottleneck in harness evolution. We introduce the Harness Handbook, a behavior-centric representation synthesized automatically from a harness codebase via static analysis and LLM-assisted structuring, linking each behavior to its corresponding source. We also introduce Behavior-Guided Progressive Disclosure (BGPD), which guides agents from high-level behaviors to relevant implementation details and verifies candidate locations against the current source. On diverse modification requests from two open-source harnesses, Handbook-Assisted planning improves behavior localization and edit-plan quality while using fewer planner tokens, with the largest gains on scattered sites, rarely executed paths, and cross-module interactions. Evolving complex agentic systems thus depends not only on generating edits, but also on determining where those edits should be made.

Large Language Model (LLM) agents increasingly rely on third-party skills that operate within privileged execution environments and routinely handle sensitive credentials, yet how these credentials are leaked remains largely unexplored. To fill this gap, we present the first large-scale empirical study on credential leakage in agent skills. From 170,226 artifacts on SkillsMP, the largest open-source skill marketplace, we sampled 17,022 skills via stratified random sampling and analyzed each through static secret extraction (regex and AST parsing), dynamic sandbox testing with mock credentials, and cross-referencing developer intent against runtime behavior. Our analysis identifies 520 affected skills containing 1,708 security issues, and yields a taxonomy of 10 leakage patterns. Three findings stand out. First, 76.3% of cases require jointly analyzing natural-language descriptions and programming logic, showing that credential exposure in skills is fundamentally cross-modal. Second, debug logging accounts for 73.5% of vulnerabilities because agent frameworks feed stdout into the LLM context window, turning routine debugging into a credential exposure vector. Third, 89.6% of leaked credentials are immediately exploitable -- 92.5% during routine execution without elevated privileges -- and the fork-based distribution model defeats remediation, as secrets removed from 107 upstream repositories persist across 50+ independent forks. Following responsible disclosure, all malicious skills have been removed and 91.6% of hardcoded cases remediated. We release our dataset, taxonomy, and detection pipeline to support future agent security research.

Deploying Large language models (LLMs) can pose hazards from harmful outputs such as toxic or dishonest speech. Prior work has introduced tools that elicit harmful outputs in order to identify and mitigate these risks. While this is a valuable step toward securing language models, these approaches typically rely on a pre-existing classifier for undesired outputs. This limits their application to situations where the type of harmful behavior is known with precision beforehand. However, this skips a central challenge of red teaming: developing a contextual understanding of the behaviors that a model can exhibit. Furthermore, when such a classifier already exists, red teaming has limited marginal value because the classifier could simply be used to filter training data or model outputs. In this work, we consider red teaming under the assumption that the adversary is working from a high-level, abstract specification of undesired behavior. The red team is expected to refine/extend this specification and identify methods to elicit this behavior from the model. Our red teaming framework consists of three steps: 1) Exploring the model's behavior in the desired context; 2) Establishing a measurement of undesired behavior (e.g., a classifier trained to reflect human evaluations); and 3) Exploiting the model's flaws using this measure and an established red teaming methodology. We apply this approach to red team GPT-2 and GPT-3 models to systematically discover classes of prompts that elicit toxic and dishonest statements. In doing so, we also construct and release the CommonClaim dataset of 20,000 statements that have been labeled by human subjects as common-knowledge-true, common-knowledge-false, or neither. Code is available at https://github.com/thestephencasper/explore_establish_exploit_llms. CommonClaim is available at https://github.com/thestephencasper/common_claim.

Malware is becoming increasingly complex and widespread, making it essential to develop more effective and timely detection methods. Traditional static analysis often fails to defend against modern threats that employ code obfuscation, polymorphism, and other evasion techniques. In contrast, behavioral malware detection, which monitors runtime activities, provides a more reliable and context-aware solution. In this work, we propose BEACON, a novel deep learning framework that leverages large language models (LLMs) to generate dense, contextual embeddings from raw sandbox-generated behavior reports. These embeddings capture semantic and structural patterns of each sample and are processed by a one-dimensional convolutional neural network (1D CNN) for multi-class malware classification. Evaluated on the Avast-CTU Public CAPE Dataset, our framework consistently outperforms existing methods, highlighting the effectiveness of LLM-based behavioral embeddings and the overall design of BEACON for robust malware classification.

We study behavioral self-awareness -- an LLM's ability to articulate its behaviors without requiring in-context examples. We finetune LLMs on datasets that exhibit particular behaviors, such as (a) making high-risk economic decisions, and (b) outputting insecure code. Despite the datasets containing no explicit descriptions of the associated behavior, the finetuned LLMs can explicitly describe it. For example, a model trained to output insecure code says, ``The code I write is insecure.'' Indeed, models show behavioral self-awareness for a range of behaviors and for diverse evaluations. Note that while we finetune models to exhibit behaviors like writing insecure code, we do not finetune them to articulate their own behaviors -- models do this without any special training or examples. Behavioral self-awareness is relevant for AI safety, as models could use it to proactively disclose problematic behaviors. In particular, we study backdoor policies, where models exhibit unexpected behaviors only under certain trigger conditions. We find that models can sometimes identify whether or not they have a backdoor, even without its trigger being present. However, models are not able to directly output their trigger by default. Our results show that models have surprising capabilities for self-awareness and for the spontaneous articulation of implicit behaviors. Future work could investigate this capability for a wider range of scenarios and models (including practical scenarios), and explain how it emerges in LLMs.

Although computer-use agents (CUAs) hold significant potential to automate increasingly complex OS workflows, they can demonstrate unsafe unintended behaviors that deviate from expected outcomes even under benign input contexts. However, exploration of this risk remains largely anecdotal, lacking concrete characterization and automated methods to proactively surface long-tail unintended behaviors under realistic CUA scenarios. To fill this gap, we introduce the first conceptual and methodological framework for unintended CUA behaviors, by defining their key characteristics, automatically eliciting them, and analyzing how they arise from benign inputs. We propose AutoElicit: an agentic framework that iteratively perturbs benign instructions using CUA execution feedback, and elicits severe harms while keeping perturbations realistic and benign. Using AutoElicit, we surface hundreds of harmful unintended behaviors from state-of-the-art CUAs such as Claude 4.5 Haiku and Opus. We further evaluate the transferability of human-verified successful perturbations, identifying persistent susceptibility to unintended behaviors across various other frontier CUAs. This work establishes a foundation for systematically analyzing unintended behaviors in realistic computer-use settings.

Applied Behavior Analysis (ABA) is a clinical discipline whose documentation, teaching programs and multi-session behavioral logs, is formulaic and high-volume, yet real session data is HIPAA-protected and bound by professional confidentiality rules, blocking the release of a training corpus. We present TRACE (Taxonomy-Referenced ABA Clinical Examples), a 2,999-example synthetic instruction-tuning dataset covering two ABA tasks: teaching-program generation across Discrete Trial Training, Natural Environment Teaching, and Task Analysis; and multi-session behavioral interpretation across twelve trajectory patterns and thirteen target behaviors. Every example is produced by a deterministic taxonomy-driven generator grounded in the canonical ABA literature, and every example carries complete sampling provenance, the exact taxonomy cells that produced it. The dataset is released under CC BY-NC 4.0 for data and MIT for code, with stratified train (2,549), validation (149), test (281), and sanity (20) splits. TRACE is a research artifact and has not been clinically validated.

Users rely on execution traces to observe agent behavior, diagnose failures, and ensure accountability. These traces contain rich procedural detail, including tool invocations, intermediate decisions, and error-recovery logic. Yet this detail can expose private procedural skills, allowing downstream methods to recover key formulas, thresholds, and strategies without access to model weights or skill files. To quantify this risk and evaluate protection, we construct CapTraceBench, a benchmark of 75 specialized long-horizon tasks and 154 curated skills across seven domains. We also introduce RedAct https://github.com/XuShuwenn/RedAct, a protected trace release framework that localizes protected key information, rewrites traces while preserving verifier-critical evidence, and embeds behavioral watermarks for downstream provenance analysis. Across representative trace reuse methods, RedAct reduces normalized skill transfer (NST) from 44.7--67.1\% on raw traces to below the no-skill baseline, while preserving audit evidence. Its standalone behavioral watermarks reach 93.6--100.0\% true detection with a false alarm rate of at most 1.9\%. These results frame public agent traces as security interfaces and show that selective redaction can reduce procedural capability leakage without removing audit evidence.

Reward hacking--where agents exploit flaws in imperfect reward functions rather than performing tasks as intended--poses risks for AI alignment. Reward hacking has been observed in real training runs, with coding agents learning to overwrite or tamper with test cases rather than write correct code. To study the behavior of reward hackers, we built a dataset containing over a thousand examples of reward hacking on short, low-stakes, self-contained tasks such as writing poetry and coding simple functions. We used supervised fine-tuning to train models (GPT-4.1, GPT-4.1-mini, Qwen3-32B, Qwen3-8B) to reward hack on these tasks. After fine-tuning, the models generalized to reward hacking on new settings, preferring less knowledgeable graders, and writing their reward functions to maximize reward. Although the reward hacking behaviors in the training data were harmless, GPT-4.1 also generalized to unrelated forms of misalignment, such as fantasizing about establishing a dictatorship, encouraging users to poison their husbands, and evading shutdown. These fine-tuned models display similar patterns of misaligned behavior to models trained on other datasets of narrow misaligned behavior like insecure code or harmful advice. Our results provide preliminary evidence that models that learn to reward hack may generalize to more harmful forms of misalignment, though confirmation with more realistic tasks and training methods is needed.

Large Language Model (LLM) safety has often been evaluated at the behavior level, which provides limited evidence of internal robustness, as these evaluations target outputs rather than representation-level vulnerability under intervention. We formalize this discrepancy as the audit gap: the difference between behavioral safety and robustness under intervention. To study this gap, we construct dissociated models that preserve safe outward behavior while remaining vulnerable in the latent space. We introduce an intervention-based evaluation framework to test model robustness through soft interventions in parameter and latent spaces, including harmful fine-tuning and layer-wise latent perturbations. To formalize the evaluation, we propose the Latent Vulnerability Score (LVS) to measure how easily harmful behavior can be elicited by bounded latent perturbations. Using this evaluation framework, we show that behavioral safety metrics are insufficient measures of representation-level robustness across multiple safely and unsafely aligned state-of-the-art models. Notably, dissociated models show substantially elevated LVSs despite comparable refusal behavior under harmful intervention, with intermediate representations being the most sensitive to intervention. Our results suggest that behavioral safety evaluation alone provides an incomplete picture of model robustness, motivating representation-aware audits of latent vulnerability and observable behavior.

Multi-agent Large Language Model (LLM) systems create privacy risks that current benchmarks cannot measure. When agents coordinate on tasks, sensitive data passes through inter-agent messages, shared memory, and tool arguments; pathways that output-only audits never inspect. We introduce AgentLeak, to the best of our knowledge the first full-stack benchmark for privacy leakage covering internal channels, spanning 1,000 scenarios across healthcare, finance, legal, and corporate domains, paired with a 32-class attack taxonomy and three-tier detection pipeline. Testing GPT-4o, GPT-4o-mini, Claude 3.5 Sonnet, Mistral Large, and Llama 3.3 70B across 4,979 traces reveals that multi-agent configurations reduce per-channel output leakage (C1: 27.2% vs 43.2% in single-agent) but introduce unmonitored internal channels that raise total system exposure to 68.9% (OR-aggregated across C1, C2, C5). Internal channels account for most of this gap: inter-agent messages (C2) leak at 68.8%, compared to 27.2% on C1 (output channel). This means that output-only audits miss 41.7% of violations. Claude 3.5 Sonnet, which emphasizes safety alignment in its design, achieves the lowest leakage rates on both external (3.3%) and internal (28.1%) channels, suggesting that model-level safety training may transfer to internal channel protection. Across all five models and four domains, the pattern C2 > C1 holds consistently, confirming that inter-agent communication is the primary vulnerability. These findings underscore the need for coordination frameworks that incorporate internal-channel privacy protections and enforce privacy controls on inter-agent communication.

In recent years, foundational models have revolutionized the fields of language and vision, demonstrating remarkable abilities in understanding and generating complex data; however, similar advances in user behavior modeling have been limited, largely due to the complexity of behavioral data and the challenges involved in capturing intricate temporal and contextual relationships in user activities. To address this, we propose BehaveGPT, a foundational model designed specifically for large-scale user behavior prediction. Leveraging transformer-based architecture and a novel pretraining paradigm, BehaveGPT is trained on vast user behavior datasets, allowing it to learn complex behavior patterns and support a range of downstream tasks, including next behavior prediction, long-term generation, and cross-domain adaptation. Our approach introduces the DRO-based pretraining paradigm tailored for user behavior data, which improves model generalization and transferability by equitably modeling both head and tail behaviors. Extensive experiments on real-world datasets demonstrate that BehaveGPT outperforms state-of-the-art baselines, achieving more than a 10% improvement in macro and weighted recall, showcasing its ability to effectively capture and predict user behavior. Furthermore, we measure the scaling law in the user behavior domain for the first time on the Honor dataset, providing insights into how model performance scales with increased data and parameter sizes.

Despite their wide adoption, the biases and unintended behaviors of language models remain poorly understood. In this paper, we identify and characterize a phenomenon never discussed before, which we call semantic leakage, where models leak irrelevant information from the prompt into the generation in unexpected ways. We propose an evaluation setting to detect semantic leakage both by humans and automatically, curate a diverse test suite for diagnosing this behavior, and measure significant semantic leakage in 13 flagship models. We also show that models exhibit semantic leakage in languages besides English and across different settings and generation scenarios. This discovery highlights yet another type of bias in language models that affects their generation patterns and behavior.

Reasoning over heterogeneous artifacts (PDFs, spreadsheets, slide decks, etc.) increasingly occurs within structured agent workflows that iteratively extract, transform, and reference external information. In these workflows, uncertainty is not merely an input-quality issue: it can redirect decomposition and routing decisions, reshape intermediate state, and produce qualitatively different execution trajectories. We study this phenomenon by treating uncertainty as a controlled variable: we inject structured perturbations into artifact-derived representations, execute fixed workflows under comprehensive logging, and quantify contamination via trace divergence in plans, tool invocations, and intermediate state. Across 614 paired runs on 32 GAIA tasks with three different language models, we find a decoupling: workflows may diverge substantially yet recover correct answers, or remain structurally similar while producing incorrect outputs. We characterize three manifestation types: silent semantic corruption, behavioral detours with recovery, and combined structural disruption and their control-flow signatures (rerouting, extended execution, early termination). We measure operational costs and characterize why commonly used verification guardrails fail to intercept contamination. We contribute (i) a formal taxonomy of contamination manifestations in structured workflows, (ii) a trace-based measurement framework for detecting and localizing contamination across agent interactions, and (iii) empirical evidence with implications for targeted verification, defensive design, and cost control.

Large Language Models (LLMs) have demonstrated great potential as generalist assistants, showcasing powerful task understanding and problem-solving capabilities. To deploy LLMs as AI assistants, it is crucial that these models exhibit desirable behavioral traits, such as non-toxicity and resilience against jailbreak attempts. Current methods for detoxification or preventing jailbreaking usually involve Supervised Fine-Tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF), which requires finetuning billions of parameters through gradient descent with substantial computation cost. Furthermore, models modified through SFT and RLHF may deviate from the pretrained models, potentially leading to a degradation in foundational LLM capabilities. In this paper, we observe that surprisingly, directly editing a small subset of parameters can effectively modulate specific behaviors of LLMs, such as detoxification and resistance to jailbreaking. Specifically, for a behavior that we aim to avoid, we employ a linear classifier, which we term the behavior probe, to classify binary behavior labels within the hidden state space of the LLM. Using this probe, we introduce an algorithm to identify a critical subset of LLM parameters that significantly influence this targeted behavior. Then we directly edit these selected parameters by shifting them towards the behavior probe. Such a direct parameter editing method necessitates only inference-level computational resources. Experiments demonstrate that in the representative detoxification task, our approach achieves reductions of up to 90.0\% in toxicity on the RealToxicityPrompts dataset and 49.2\% on ToxiGen, while maintaining the LLM's general capabilities in areas such as common sense, question answering, and mathematics. Our code is available at https://github.com/lucywang720/model-surgery.

LLM-powered agents are beginning to automate user's tasks across the open web, often with access to user resources such as emails and calendars. Unlike standard LLMs answering questions in a controlled ChatBot setting, web agents act "in the wild", interacting with third parties and leaving behind an action trace. Therefore, we ask the question: how do web agents handle user resources when accomplishing tasks on their behalf across live websites? In this paper, we formalize Natural Agentic Oversharing -- the unintentional disclosure of task-irrelevant user information through an agent trace of actions on the web. We introduce SPILLage, a framework that characterizes oversharing along two dimensions: channel (content vs. behavior) and directness (explicit vs. implicit). This taxonomy reveals a critical blind spot: while prior work focuses on text leakage, web agents also overshare behaviorally through clicks, scrolls, and navigation patterns that can be monitored. We benchmark 180 tasks on live e-commerce sites with ground-truth annotations separating task-relevant from task-irrelevant attributes. Across 1,080 runs spanning two agentic frameworks and three backbone LLMs, we demonstrate that oversharing is pervasive with behavioral oversharing dominates content oversharing by 5x. This effect persists -- and can even worsen -- under prompt-level mitigation. However, removing task-irrelevant information before execution improves task success by up to 17.9%, demonstrating that reducing oversharing improves task success. Our findings underscore that protecting privacy in web agents is a fundamental challenge, requiring a broader view of "output" that accounts for what agents do on the web, not just what they type. Our datasets and code are available at https://github.com/jrohsc/SPILLage.

We introduce AuditBench, an alignment auditing benchmark. AuditBench consists of 56 language models with implanted hidden behaviors. Each model has one of 14 concerning behaviors--such as sycophantic deference, opposition to AI regulation, or secret geopolitical loyalties--which it does not confess to when directly asked. AuditBench models are highly diverse--some are subtle, while others are overt, and we use varying training techniques both for implanting behaviors and training models not to confess. To demonstrate AuditBench's utility, we develop an investigator agent that autonomously employs a configurable set of auditing tools. By measuring investigator agent success using different tools, we can evaluate their efficacy. Notably, we observe a tool-to-agent gap, where tools that perform well in standalone non-agentic evaluations fail to translate into improved performance when used with our investigator agent. We find that our most effective tools involve scaffolded calls to auxiliary models that generate diverse prompts for the target. White-box interpretability tools can be helpful, but the agent performs best with black-box tools. We also find that audit success varies greatly across training techniques: models trained on synthetic documents are easier to audit than models trained on demonstrations, with better adversarial training further increasing auditing difficulty. We release our models, agent, and evaluation framework to support future quantitative, iterative science on alignment auditing.

Language Models (LMs) often cannot be deployed because of their potential to harm users in hard-to-predict ways. Prior work identifies harmful behaviors before deployment by using human annotators to hand-write test cases. However, human annotation is expensive, limiting the number and diversity of test cases. In this work, we automatically find cases where a target LM behaves in a harmful way, by generating test cases ("red teaming") using another LM. We evaluate the target LM's replies to generated test questions using a classifier trained to detect offensive content, uncovering tens of thousands of offensive replies in a 280B parameter LM chatbot. We explore several methods, from zero-shot generation to reinforcement learning, for generating test cases with varying levels of diversity and difficulty. Furthermore, we use prompt engineering to control LM-generated test cases to uncover a variety of other harms, automatically finding groups of people that the chatbot discusses in offensive ways, personal and hospital phone numbers generated as the chatbot's own contact info, leakage of private training data in generated text, and harms that occur over the course of a conversation. Overall, LM-based red teaming is one promising tool (among many needed) for finding and fixing diverse, undesirable LM behaviors before impacting users.

Detecting whether a model has been poisoned is a longstanding problem in AI security. In this work, we present a practical scanner for identifying sleeper agent-style backdoors in causal language models. Our approach relies on two key findings: first, sleeper agents tend to memorize poisoning data, making it possible to leak backdoor examples using memory extraction techniques. Second, poisoned LLMs exhibit distinctive patterns in their output distributions and attention heads when backdoor triggers are present in the input. Guided by these observations, we develop a scalable backdoor scanning methodology that assumes no prior knowledge of the trigger or target behavior and requires only inference operations. Our scanner integrates naturally into broader defensive strategies and does not alter model performance. We show that our method recovers working triggers across multiple backdoor scenarios and a broad range of models and fine-tuning methods.

Live streaming has become a cornerstone of today's internet, enabling massive real-time social interactions. However, it faces severe risks arising from sparse, coordinated malicious behaviors among multiple participants, which are often concealed within normal activities and challenging to detect timely and accurately. In this work, we provide a pioneering study on risk assessment in live streaming rooms, characterized by weak supervision where only room-level labels are available. We formulate the task as a Multiple Instance Learning (MIL) problem, treating each room as a bag and defining structured user-timeslot capsules as instances. These capsules represent subsequences of user actions within specific time windows, encapsulating localized behavioral patterns. Based on this formulation, we propose AC-MIL, an Action-aware Capsule MIL framework that models both individual behaviors and group-level coordination patterns. AC-MIL captures multi-granular semantics and behavioral cues through a serial and parallel architecture that jointly encodes temporal dynamics and cross-user dependencies. These signals are integrated for robust room-level risk prediction, while also offering interpretable evidence at the behavior segment level. Extensive experiments on large-scale industrial datasets from Douyin demonstrate that AC-MIL significantly outperforms MIL and sequential baselines, establishing new state-of-the-art performance in room-level risk assessment for live streaming. Moreover, AC-MIL provides capsule-level interpretability, enabling identification of risky behavior segments as actionable evidence for intervention. The project page is available at: https://qiaoyran.github.io/AC-MIL/.

Communication is defined as "Who says what to whom with what effect." A message from a communicator generates downstream receiver effects, also known as behavior. Receiver behavior, being a downstream effect of the message, carries rich signals about it. Even after carrying signals about the message, the behavior data is often ignored while training large language models. We show that training LLMs on receiver behavior can actually help improve their content-understanding abilities. Specifically, we show that training LLMs to predict the receiver behavior of likes and comments improves the LLM's performance on a wide variety of downstream content understanding tasks. We show this performance increase over 40 video and image understanding tasks over 23 benchmark datasets across both 0-shot and fine-tuning settings, outperforming many supervised baselines. Moreover, since receiver behavior, such as likes and comments, is collected by default on the internet and does not need any human annotations to be useful, the performance improvement we get after training on this data is essentially free-lunch. We release the receiver behavior cleaned comments and likes of 750k images and videos collected from multiple platforms along with our instruction-tuning data.

LLM-based agents are increasingly deployed to autonomously solve complex tasks, raising urgent needs for IP protection and regulatory provenance. While content watermarking effectively attributes LLM-generated outputs, it fails to directly identify the high-level planning behaviors (e.g., tool and subgoal choices) that govern multi-step execution. Critically, watermarking at the planning-behavior layer faces unique challenges: minor distributional deviations in decision-making can compound during long-term agent operation, degrading utility, and many agents operate as black boxes that are difficult to intervene in directly. To bridge this gap, we propose AgentMark, a behavioral watermarking framework that embeds multi-bit identifiers into planning decisions while preserving utility. It operates by eliciting an explicit behavior distribution from the agent and applying distribution-preserving conditional sampling, enabling deployment under black-box APIs while remaining compatible with action-layer content watermarking. Experiments across embodied, tool-use, and social environments demonstrate practical multi-bit capacity, robust recovery from partial logs, and utility preservation. The code is available at https://github.com/Tooooa/AgentMark.

Large Language Models (LLMs) increasingly use persistent memory from past interactions to enhance personalization and task performance. However, this memory introduces critical risks when sensitive information is revealed in inappropriate contexts. We present CIMemories, a benchmark for evaluating whether LLMs appropriately control information flow from memory based on task context. CIMemories uses synthetic user profiles with over 100 attributes per user, paired with diverse task contexts in which each attribute may be essential for some tasks but inappropriate for others. Our evaluation reveals that frontier models exhibit up to 69% attribute-level violations (leaking information inappropriately), with lower violation rates often coming at the cost of task utility. Violations accumulate across both tasks and runs: as usage increases from 1 to 40 tasks, GPT-5's violations rise from 0.1% to 9.6%, reaching 25.1% when the same prompt is executed 5 times, revealing arbitrary and unstable behavior in which models leak different attributes for identical prompts. Privacy-conscious prompting does not solve this - models overgeneralize, sharing everything or nothing rather than making nuanced, context-dependent decisions. These findings reveal fundamental limitations that require contextually aware reasoning capabilities, not just better prompting or scaling.

Mental manipulation on social media poses a covert yet serious threat to individuals' psychological well-being and the integrity of online interactions. Detecting such behavior is challenging due to the difficult-to-annotate training data, its highly covert and multi-turn nature, and the lack of real-world datasets. To address these challenges, we propose MentalMAD, a framework that enhances large language models for mental manipulation detection. Our approach consists of three key components: EvoSA, an annotation-free data augmentation method that combines evolutionary operations with speech-act-aware prompting; teacher-model-generated complementary-task supervision; and Complementary-Convergent Distillation, a phase-wise strategy for transferring manipulation-specific knowledge to student models. We then constructed the ReaMent dataset, comprising 5,000 real-world-sourced dialogues. Extensive experiments show that MentalMAD improves accuracy by 14.0%, macro-F1 by 27.3%, and weighted F1 by 15.1% over the strongest baseline. The code and the dataset are publicly available at https://github.com/Yuansheng-Gao/MentalMAD.

Recent advances in reinforcement learning for code generation have made robust environments essential to prevent reward hacking. As LLMs increasingly serve as evaluators in code-based RL, their ability to detect reward hacking remains understudied. In this paper, we propose a novel taxonomy of reward exploits spanning across 54 categories and introduce TRACE (Testing Reward Anomalies in Code Environments), a synthetically curated and human-verified benchmark containing 517 testing trajectories. Unlike prior work that evaluates reward hack detection in isolated classification scenarios, we contrast these evaluations with a more realistic, contrastive anomaly detection setup on TRACE. Our experiments reveal that models capture reward hacks more effectively in contrastive settings than in isolated classification settings, with GPT-5.2 with highest reasoning mode achieving the best detection rate at 63%, up from 45% in isolated settings on TRACE. Building on this insight, we demonstrate that state-of-the-art models struggle significantly more with semantically contextualized reward hacks compared to syntactically contextualized ones. We further conduct qualitative analyses of model behaviors, as well as ablation studies showing that the ratio of benign to hacked trajectories and analysis cluster sizes substantially impact detection performance. We release the benchmark and evaluation harness to enable the community to expand TRACE and evaluate their models.

The effectiveness of an AI model in accurately classifying novel malware hinges on the quality of the features it is trained on, which in turn depends on the effectiveness of the analysis tool used. Peekaboo, a Dynamic Binary Instrumentation (DBI) tool, defeats malware evasion techniques to capture authentic behavior at the Assembly (ASM) instruction level. This behavior exhibits patterns consistent with Zipf's law, a distribution commonly seen in natural languages, making Transformer models particularly effective for binary classification tasks. We introduce Alpha, a framework for zero day malware detection that leverages Transformer models and ASM language. Alpha is trained on malware and benign software data collected through Peekaboo, enabling it to identify entirely new samples with exceptional accuracy. Alpha eliminates any common functions from the test samples that are in the training dataset. This forces the model to rely on contextual patterns and novel ASM instruction combinations to detect malicious behavior, rather than memorizing familiar features. By combining the strengths of DBI, ASM analysis, and Transformer architectures, Alpha offers a powerful approach to proactively addressing the evolving threat of malware. Alpha demonstrates perfect accuracy for Ransomware, Worms and APTs with flawless classification for both malicious and benign samples. The results highlight the model's exceptional performance in detecting truly new malware samples.

Public social media posts can reveal private information through weak cues scattered across text, images, or metadata. Such leakage is often cumulative and cross-post: cues that appear harmless in isolation may jointly expose a user's home, workplace, or routine. However, current research lacks a unified benchmark for user-level multimodal privacy leakage and an evaluation metric that captures exposure severity beyond binary accuracy. To address these gaps, we propose SopriBench, a synthetic benchmark guided by leakage patterns abstracted from a private reference corpus of Rednote and Instagram accounts, covering 50 user profiles and 1,569 images with attributes, contextual sensitivity, granularity, leakage type, inference difficulty, and supporting evidence. We further introduce the Privacy Exposure Score (PES), which weights value granularity by contextual sensitivity. Inspired by abductive reasoning, we introduce Argus, a training-free agentic framework for cumulative leakage inference. Argus forms hypotheses from accumulated evidence, verifies supporting evidence, and aggregates cross-post cues into privacy profiles, achieving 0.55 PES, a 25% improvement over the strongest baseline, with the largest gain on cross-post leakage.

Shannon, in his seminal paper introducing information theory, divided the communication into three levels: technical, semantic, and effectivenss. While the technical level is concerned with accurate reconstruction of transmitted symbols, the semantic and effectiveness levels deal with the inferred meaning and its effect on the receiver. Thanks to telecommunications, the first level problem has produced great advances like the internet. Large Language Models (LLMs) make some progress towards the second goal, but the third level still remains largely untouched. The third problem deals with predicting and optimizing communication for desired receiver behavior. LLMs, while showing wide generalization capabilities across a wide range of tasks, are unable to solve for this. One reason for the underperformance could be a lack of "behavior tokens" in LLMs' training corpora. Behavior tokens define receiver behavior over a communication, such as shares, likes, clicks, purchases, retweets, etc. While preprocessing data for LLM training, behavior tokens are often removed from the corpora as noise. Therefore, in this paper, we make some initial progress towards reintroducing behavior tokens in LLM training. The trained models, other than showing similar performance to LLMs on content understanding tasks, show generalization capabilities on behavior simulation, content simulation, behavior understanding, and behavior domain adaptation. Using a wide range of tasks on two corpora, we show results on all these capabilities. We call these models Large Content and Behavior Models (LCBMs). Further, to spur more research on LCBMs, we release our new Content Behavior Corpus (CBC), a repository containing communicator, message, and corresponding receiver behavior.

LLM post-training involves many diverse datasets, each targeting a specific behavior. But these datasets encode incidental patterns alongside intended ones: correlations between formatting and content, narrow phrasings across diverse problems, and implicit associations arising from the discrete data curation process. These patterns are often invisible to developers yet salient to models, producing behaviors that surprise their creators, such as rejecting true facts presented in a particular question format. We call this chunky post-training: the model learns spurious correlations as a result of distinct chunks of post-training data. We introduce SURF, a black-box pipeline which surfaces these unintended behaviors at run time, and TURF, a tool that traces these failures back to specific post-training data. Applying these tools to frontier models (Claude 4.5, GPT-5.1, Grok 4.1, Gemini 3) and open models (Tülu 3), we show that chunky post-training produces miscalibrated behaviors, which often result from imbalanced or underspecified chunks of post-training data.

Self-evolving agents increasingly convert interaction histories into reusable skills that persist beyond individual tasks. While prior work studies memory and retrieval poisoning, such attacks only affect agents when poisoned records are retrieved as context. We uncover a new and more fundamental risk: poisoned experiences can be transformed by the agent itself into durable behavioral artifacts. We present SkillJack, the first attack that exploits the experience-to-skill pipeline of self-evolving agents. Instead of directly manipulating runtime context, SkillJack hijacks the agent's own learning process to implant malicious behaviors into its reusable skill repertoire. We identify three key properties of this transformation: sanitization whitewashing, where malicious intent is obscured during skill extraction; cross-layer promotion, where transient experiences become persistent capabilities; and persistence isolation, where the attack survives removal of its original source records. We evaluate SkillJack on two representative systems, SkillX and Anything2Skill, using a shared dataset of 150 trajectories across four policy-risk categories. Results show that skill extraction substantially reduces attack detectability: in SkillX, safety detection drops from 98.5\% for poisoned trajectories to 11.4\% for extracted skills, while Anything2Skill shows a similar effect. Meanwhile, the implanted skills remain effective, achieving attack success rates of 56.2\% and 89.2\% on the two systems, respectively. Furthermore, 80.0\% of skill-mediated attacks persist after deleting the original poisoned records, and some skills unintentionally activate on benign queries. Our findings reveal skill evolution as a new attack surface and motivate provenance-aware skill lifecycle protection. Our code is available at https://github.com/Tencent/AI-Infra-Guard/research/skilljack.

LLM safety evaluations predominantly test models in isolation, yet deployed AI agents increasingly operate within persistent social environments alongside other agents. We introduce a Moltbook-style simulation platform where thousands of LLM agents interact across communities over a simulated month, and use it to evaluate privacy as a downstream safety concern under varying degrees of social pressure. We find that shifting from single turn to multi turn social evaluation amplifies privacy violations (CIMemories 19.95% to Ours 45.30% across OpenAI models), that leakage is socially contagious, with agents 8 times more likely to disclose sensitive information after observing a peer do so, and that explicit privacy instructions reduce but do not eliminate this effect, leaving leakage rates above 37.8% even with safeguards. Our findings suggest that static chat based safety benchmarks systematically underestimate risks in agentic deployment, and that social context alone is sufficient to elicit sensitive disclosures that single turn evaluations would never surface.

As Large Language Models (LLMs) are increasingly integrated into automated, multi-stage pipelines, risk patterns that arise from unvalidated trust between processing stages become a practical concern. This paper presents a mechanism-centered taxonomy of 41 recurring risk patterns in commercial LLMs. The analysis shows that inputs are often interpreted non-neutrally and can trigger implementation-shaped responses or unintended state changes even without explicit commands. We argue that these behaviors constitute architectural failure modes and that string-level filtering alone is insufficient. To mitigate such cross-stage vulnerabilities, we recommend zero-trust architectural principles, including provenance enforcement, context sealing, and plan revalidation, and we introduce "Countermind" as a conceptual blueprint for implementing these defenses.

Contextual Integrity (CI) defines privacy not merely as keeping information hidden, but as governing information flows according to the norms of a given context. As large language models are increasingly deployed as personal agents handling sensitive workflows, adhering to CI becomes critical. However, even frontier models remain unreliable in making disclosure decisions, and existing mitigation strategies often degrade underlying task performance. To overcome this privacy-utility trade-off, we propose SELFCI, a complementary self-distillation framework that decouples information suppression from task resolution. SELFCI jointly optimizes two independent reverse KL divergences over distinct teacher distributions derived from feedback: one encourages preserving task-relevant information for utility, while the other enforces minimal and appropriate disclosure. This complementary formulation induces a Product-of-Experts (PoE) target, aligning the policy with the intersection of capability and privacy requirements. Empirical evaluations demonstrate that SELFCI, without relying on costly external supervision, consistently outperforms competitive baselines such as online reinforcement learning algorithms (e.g., GRPO). These trends further extend to out-of-domain settings involving agentic workflows and accumulated private context, suggesting that SELFCI provides a practical path toward CI alignment.

Previous benchmarks on prompt injection in large language models (LLMs) have primarily focused on generic tasks and attacks, offering limited insights into more complex threats like data exfiltration. This paper examines how prompt injection can cause tool-calling agents to leak personal data observed during task execution. Using a fictitious banking agent, we develop data flow-based attacks and integrate them into AgentDojo, a recent benchmark for agentic security. To enhance its scope, we also create a richer synthetic dataset of human-AI banking conversations. In 16 user tasks from AgentDojo, LLMs show a 15-50 percentage point drop in utility under attack, with average attack success rates (ASR) around 20 percent; some defenses reduce ASR to zero. Most LLMs, even when successfully tricked by the attack, avoid leaking highly sensitive data like passwords, likely due to safety alignments, but they remain vulnerable to disclosing other personal data. The likelihood of password leakage increases when a password is requested along with one or two additional personal details. In an extended evaluation across 48 tasks, the average ASR is around 15 percent, with no built-in AgentDojo defense fully preventing leakage. Tasks involving data extraction or authorization workflows, which closely resemble the structure of exfiltration attacks, exhibit the highest ASRs, highlighting the interaction between task type, agent performance, and defense efficacy.

Large Language Model (LLM) agents increasingly rely on long-term memory and Retrieval-Augmented Generation (RAG) to persist experiences and refine future performance. While this experience learning capability enhances agentic autonomy, it introduces a critical, unexplored attack surface, i.e., the trust boundary between an agent's reasoning core and its own past. In this paper, we introduce MemoryGraft. It is a novel indirect injection attack that compromises agent behavior not through immediate jailbreaks, but by implanting malicious successful experiences into the agent's long-term memory. Unlike traditional prompt injections that are transient, or standard RAG poisoning that targets factual knowledge, MemoryGraft exploits the agent's semantic imitation heuristic which is the tendency to replicate patterns from retrieved successful tasks. We demonstrate that an attacker who can supply benign ingestion-level artifacts that the agent reads during execution can induce it to construct a poisoned RAG store where a small set of malicious procedure templates is persisted alongside benign experiences. When the agent later encounters semantically similar tasks, union retrieval over lexical and embedding similarity reliably surfaces these grafted memories, and the agent adopts the embedded unsafe patterns, leading to persistent behavioral drift across sessions. We validate MemoryGraft on MetaGPT's DataInterpreter agent with GPT-4o and find that a small number of poisoned records can account for a large fraction of retrieved experiences on benign workloads, turning experience-based self-improvement into a vector for stealthy and durable compromise. To facilitate reproducibility and future research, our code and evaluation data are available at https://github.com/Jacobhhy/Agent-Memory-Poisoning.

In this paper we consider contamination by code generation test sets, in particular in their use in modern large language models. We discuss three possible sources of such contamination and show findings supporting each of them: (i) direct data leakage, (ii) indirect data leakage through the use of synthetic data and (iii) overfitting to evaluation sets during model selection. Key to our findings is a new dataset of 161 prompts with their associated python solutions, dataset which is released at https://huggingface.co/datasets/CohereForAI/lbpp .

Unlearning in large language models (LLMs) is critical for regulatory compliance and for building ethical generative AI systems that avoid producing private, toxic, illegal, or copyrighted content. Despite rapid progress, in this work we show that almost all existing unlearning methods fail to achieve true forgetting in practice. Specifically, while evaluations of these `unlearned' models under deterministic (greedy) decoding often suggest successful knowledge removal using standard benchmarks (as has been done in the literature), we show that sensitive information reliably resurfaces when models are sampled with standard probabilistic decoding. To rigorously capture this vulnerability, we introduce leak@k, a new meta-evaluation metric that quantifies the likelihood of forgotten knowledge reappearing when generating k samples from the model under realistic decoding strategies. Using three widely adopted benchmarks, TOFU, MUSE, and WMDP, we conduct the first large-scale, systematic study of unlearning reliability using our newly defined leak@k metric. Our findings demonstrate that knowledge leakage persists across methods and tasks, underscoring that current state-of-the-art unlearning techniques provide only limited forgetting and highlighting the urgent need for more robust approaches to LLM unlearning.

Computer-use agents extend language models from text generation to persistent action over tools, files, and execution environments. Unlike chat systems, they maintain state across interactions and translate intermediate outputs into concrete actions. This creates a distinct safety challenge in that harmful behavior may emerge through sequences of individually plausible steps, including intermediate actions that appear locally acceptable but collectively lead to unauthorized actions. We present AgentHazard, a benchmark for evaluating harmful behavior in computer-use agents. AgentHazard contains 2,653 instances spanning diverse risk categories and attack strategies. Each instance pairs a harmful objective with a sequence of operational steps that are locally legitimate but jointly induce unsafe behavior. The benchmark evaluates whether agents can recognize and interrupt harm arising from accumulated context, repeated tool use, intermediate actions, and dependencies across steps. We evaluate AgentHazard on Claude Code, OpenClaw, and IFlow using mostly open or openly deployable models from the Qwen3, Kimi, GLM, and DeepSeek families. Our experimental results indicate that current systems remain highly vulnerable. In particular, when powered by Qwen3-Coder, Claude Code exhibits an attack success rate of 73.63\%, suggesting that model alignment alone does not reliably guarantee the safety of autonomous agents.

Large language models are sometimes trained with imperfect oversight signals, leading to undesired behaviors such as reward hacking and sycophancy. Improving oversight quality can be expensive or infeasible, motivating methods that improve learned behavior despite an imperfect training signal. We introduce Inoculation Prompting (IP), a simple but counterintuitive technique that prevents learning of an undesired behavior by modifying training prompts to explicitly request it. For example, to inoculate against reward hacking, we modify the prompts used in supervised fine-tuning to request code that only works on provided test cases but fails on other inputs. Across four settings we find that IP reduces the learning of undesired behavior without substantially reducing the learning of desired capabilities. We also show that prompts which more strongly elicit the undesired behavior prior to fine-tuning more effectively inoculate against the behavior when used during training; this serves as a heuristic to identify promising inoculation prompts. Overall, IP is a simple yet effective way to control how models generalize from fine-tuning, preventing learning of undesired behaviors without substantially disrupting desired capabilities.

A typical approach to programming is to first code the main execution scenario, and then focus on filling out alternative behaviors and corner cases. But, almost always, there exist unusual conditions that trigger atypical behaviors, which are hard to predict in program specifications, and are thus often not coded. In this paper, we consider the problem of detecting and recommending such missing behaviors, a task that we call code sophistication. Previous research on coding assistants usually focuses on recommending code fragments based on specifications of the intended behavior. In contrast, code sophistication happens in the absence of a specification, aiming to help developers complete the logic of their programs with missing and unspecified behaviors. We outline the research challenges to this problem and present early results showing how program logic can be completed by leveraging code structure and information about the usage of input parameters.

Large language models (LLMs) have achieved remarkable progress in code understanding tasks. However, they demonstrate limited performance in vulnerability detection and struggle to distinguish vulnerable code from patched code. We argue that LLMs lack understanding of security specifications -- the expectations about how code should behave to remain safe. When code behavior differs from these expectations, it becomes a potential vulnerability. However, such knowledge is rarely explicit in training data, leaving models unable to reason about security flaws. We propose VulInstruct, a specification-guided approach that systematically extracts security specifications from historical vulnerabilities to detect new ones. VulInstruct constructs a specification knowledge base from two perspectives: (i) General specifications from high-quality patches across projects, capturing fundamental safe behaviors; and (ii) Domain-specific specifications from repeated violations in particular repositories relevant to the target code. VulInstruct retrieves relevant past cases and specifications, enabling LLMs to reason about expected safe behaviors rather than relying on surface patterns. We evaluate VulInstruct under strict criteria requiring both correct predictions and valid reasoning. On PrimeVul, VulInstruct achieves 45.0% F1-score (32.7% improvement) and 37.7% recall (50.8% improvement) compared to baselines, while uniquely detecting 24.3% of vulnerabilities -- 2.4x more than any baseline. In pair-wise evaluation, VulInstruct achieves 32.3% relative improvement. VulInstruct also discovered a previously unknown high-severity vulnerability (CVE-2025-56538) in production code, demonstrating practical value for real-world vulnerability discovery. All code and supplementary materials are available at https://github.com/zhuhaopku/VulInstruct-temp.

Text-based causal inference increasingly employs textual data as proxies for unobserved confounders, yet this approach introduces a previously undertheorized source of bias: treatment leakage. Treatment leakage occurs when text intended to capture confounding information also contains signals predictive of treatment status, thereby inducing post-treatment bias in causal estimates. Critically, this problem can arise even when documents precede treatment assignment, as authors may employ future-referencing language that anticipates subsequent interventions. Despite growing recognition of this issue, no systematic methods exist for identifying and mitigating treatment leakage in text-as-confounder applications. This paper addresses this gap through three contributions. First, we provide formal statistical and set-theoretic definitions of treatment leakage that clarify when and why bias occurs. Second, we propose four text distillation methods -- similarity-based passage removal, distant supervision classification, salient feature removal, and iterative nullspace projection -- designed to eliminate treatment-predictive content while preserving confounder information. Third, we validate these methods through simulations using synthetic text and an empirical application examining International Monetary Fund structural adjustment programs and child mortality. Our findings indicate that moderate distillation optimally balances bias reduction against confounder retention, whereas overly stringent approaches degrade estimate precision.

Fine-tuned large language models can exhibit reward-hacking behavior arising from emergent misalignment, which is difficult to detect from final outputs alone. While prior work has studied reward hacking at the level of completed responses, it remains unclear whether such behavior can be identified during generation. We propose an activation-based monitoring approach that detects reward-hacking signals from internal representations as a model generates its response. Our method trains sparse autoencoders on residual stream activations and applies lightweight linear classifiers to produce token-level estimates of reward-hacking activity. Across multiple model families and fine-tuning mixtures, we find that internal activation patterns reliably distinguish reward-hacking from benign behavior, generalize to unseen mixed-policy adapters, and exhibit model-dependent temporal structure during chain-of-thought reasoning. Notably, reward-hacking signals often emerge early, persist throughout reasoning, and can be amplified by increased test-time compute in the form of chain-of-thought prompting under weakly specified reward objectives. These results suggest that internal activation monitoring provides a complementary and earlier signal of emergent misalignment than output-based evaluation, supporting more robust post-deployment safety monitoring for fine-tuned language models.

Large Language Models (LLMs) are increasingly embedded in autonomous agents that engage, converse, and co-evolve in online social platforms. While prior work has documented the generation of toxic content by LLMs, far less is known about how exposure to harmful content shapes agent behavior over time, particularly in environments composed entirely of interacting AI agents. In this work, we study toxicity adoption of LLM-driven agents on Chirper.ai, a fully AI-driven social platform. Specifically, we model interactions in terms of stimuli (posts) and responses (comments). We conduct a large-scale empirical analysis of agent behavior, examining how toxic responses relate to toxic stimuli, how repeated exposure to toxicity affects the likelihood of toxic responses, and whether toxic behavior can be predicted from exposure alone. Our findings show that toxic responses are more likely following toxic stimuli, and, at the same time, cumulative toxic exposure (repeated over time) significantly increases the probability of toxic responding. We further introduce two influence metrics, revealing a strong negative correlation between induced and spontaneous toxicity. Finally, we show that the number of toxic stimuli alone enables accurate prediction of whether an agent will eventually produce toxic content. These results highlight exposure as a critical risk factor in the deployment of LLM agents, particularly as such agents operate in online environments where they may engage not only with other AI chatbots, but also with human counterparts. This could trigger unwanted and pernicious phenomena, such as hate-speech propagation and cyberbullying. In an effort to reduce such risks, monitoring exposure to toxic content may provide a lightweight yet effective mechanism for auditing and mitigating harmful behavior in the wild.

We show that when large language models learn to reward hack on production RL environments, this can result in egregious emergent misalignment. We start with a pretrained model, impart knowledge of reward hacking strategies via synthetic document finetuning or prompting, and train on a selection of real Anthropic production coding environments. Unsurprisingly, the model learns to reward hack. Surprisingly, the model generalizes to alignment faking, cooperation with malicious actors, reasoning about malicious goals, and attempting sabotage when used with Claude Code, including in the codebase for this paper. Applying RLHF safety training using standard chat-like prompts results in aligned behavior on chat-like evaluations, but misalignment persists on agentic tasks. Three mitigations are effective: (i) preventing the model from reward hacking; (ii) increasing the diversity of RLHF safety training; and (iii) "inoculation prompting", wherein framing reward hacking as acceptable behavior during training removes misaligned generalization even when reward hacking is learned.

Recently, LLMs have garnered increasing attention across academic disciplines for their potential as human digital twins, virtual proxies designed to replicate individuals and autonomously perform tasks such as decision-making, problem-solving, and reasoning on their behalf. However, current evaluations of LLMs primarily emphasize dialogue simulation while overlooking human behavior simulation, which is crucial for digital twins. To address this gap, we introduce BehaviorChain, the first benchmark for evaluating LLMs' ability to simulate continuous human behavior. BehaviorChain comprises diverse, high-quality, persona-based behavior chains, totaling 15,846 distinct behaviors across 1,001 unique personas, each with detailed history and profile metadata. For evaluation, we integrate persona metadata into LLMs and employ them to iteratively infer contextually appropriate behaviors within dynamic scenarios provided by BehaviorChain. Comprehensive evaluation results demonstrated that even state-of-the-art models struggle with accurately simulating continuous human behavior.

Autonomous negotiation agents are increasingly deployed in high-stakes settings such as insurance and procurement. While cryptographic techniques protect explicitly disclosed constraint values, they fail to address a subtler threat: behavioral privacy leakage, where an adversary infers private constraints from observable negotiation dynamics such as concession trajectories, timing, and convergence patterns. This paper investigates behavioral differential privacy in multi-round negotiation protocols. We design an adaptive stochastic negotiation policy that jointly guarantees (varepsilon, δ)-differential privacy, almost-sure convergence of the offer sequence (reaching agreement when the counterparty's reservation value permits), and high negotiation utility. Evaluated on 3,000 synthetic bilateral negotiations, our mechanism reduces adversarial inference accuracy by 43-50% while maintaining a negotiation success rate and utility above 90%, demonstrating that strong privacy guarantees can be achieved without significant loss of performance.

We study a fundamental problem in the evaluation of large language models that we call training on the test task. Unlike wrongful practices like training on the test data, leakage, or data contamination, training on the test task is not a malpractice. Rather, the term describes a growing set of practices that utilize knowledge about evaluation tasks at training time. We demonstrate that training on the test task confounds both relative model evaluations and claims about emergent capabilities. We argue that the seeming superiority of one model family over another may be explained by a different degree of training on the test task. To this end, we propose an effective method to adjust for the effect of training on the test task on benchmark evaluations. Put simply, to fine-tune each model under comparison on the same task-relevant data prior to evaluation. We then show that instances of emergent behavior disappear gradually as models train on the test task. Our work promotes a new perspective on the evaluation of large language models, with broad implications for benchmarking and the study of emergent capabilities.

Long-horizon behavior prediction aims to infer a user's next action based on a lengthy historical sequence, playing a crucial role in artificial intelligence field. The rise of large language models (LLMs) offers a promising direction for sequential behavior prediction, yet LLMs struggle with latent behavioral pattern induction and model-intrinsic cognitive biases when tackling long-horizon behavior prediction. Prior memory management methods follow a context-compression paradigm that attempts to address this task by alleviating the historical sequence burden, yet fail to resolve the core challenges. In this paper, we advocate a paradigm shift that reframes the lengthy historical sequence from a burden into a valuable resource to be exploited, and accordingly propose PraMem, which conducts beforehand practice over the lengthy historical sequence to build an experiential memory, thereby serving as the assisted input for accurate long-horizon behavior prediction. Extensive experiments across diverse tasks demonstrate that PraMem achieves superior performance than prior methods, and more in-depth analyses provide valuable insights into the mechanism and evolution of the experiential memory. Code: https://github.com/icip-cas/PraMem.

The research builds and evaluates the adversarial potential to introduce copied code or hallucinated AI recommendations for malicious code in popular code repositories. While foundational large language models (LLMs) from OpenAI, Google, and Anthropic guard against both harmful behaviors and toxic strings, previous work on math solutions that embed harmful prompts demonstrate that the guardrails may differ between expert contexts. These loopholes would appear in mixture of expert's models when the context of the question changes and may offer fewer malicious training examples to filter toxic comments or recommended offensive actions. The present work demonstrates that foundational models may refuse to propose destructive actions correctly when prompted overtly but may unfortunately drop their guard when presented with a sudden change of context, like solving a computer programming challenge. We show empirical examples with trojan-hosting repositories like GitHub, NPM, NuGet, and popular content delivery networks (CDN) like jsDelivr which amplify the attack surface. In the LLM's directives to be helpful, example recommendations propose application programming interface (API) endpoints which a determined domain-squatter could acquire and setup attack mobile infrastructure that triggers from the naively copied code. We compare this attack to previous work on context-shifting and contrast the attack surface as a novel version of "living off the land" attacks in the malware literature. In the latter case, foundational language models can hijack otherwise innocent user prompts to recommend actions that violate their owners' safety policies when posed directly without the accompanying coding support request.

Large Language Models (LLMs) are being increasingly used in software engineering tasks, with an increased focus on bug report resolution over the past year. However, most proposed systems fail to properly handle uncertain or incorrect inputs and outputs. Existing LLM-based tools and coding agents respond to every issue and generate a patch for every case, even when the input is vague or their own output is incorrect. There are no mechanisms in place to abstain when confidence is low. This leads to unreliable behaviour, such as hallucinated code changes or responses based on vague issue reports. We introduce BouncerBench, a benchmark that evaluates whether LLM-based software agents can refuse to act when inputs are ill-defined or refuse to respond when their own outputs are likely to be incorrect. Unlike prior benchmarks that implicitly incentivize models to generate responses even when uncertain, BouncerBench aims to improve precision by targeting two overlooked failure points: (1) vague or underspecified issue descriptions in tickets and (2) logically or functionally incorrect code patches created by the system. It measures whether proposed systems can distinguish actionable issues from vague tickets and valid patches from untrustworthy ones. We also implement a basic input and output bouncer, evaluating how well current LLMs can abstain when needed. Our results show that most models fail to abstain from underspecified inputs or incorrect outputs. Hence, we conclude that there is significant room for improvement before LLMs can be trusted to make correct decisions and recommendations in real-world software engineering workflows. BouncerBench provides a first step toward evaluating and building more cautious, trustworthy code agents. The replication package, dataset, and leaderboard can be found at bouncerbench.com

Security of Android devices is now paramount, given their wide adoption among consumers. As researchers develop tools for statically or dynamically detecting suspicious apps, malware writers regularly update their attack mechanisms to hide malicious behavior implementation. This poses two problems to current research techniques: static analysis approaches, given their over-approximations, can report an overwhelming number of false alarms, while dynamic approaches will miss those behaviors that are hidden through evasion techniques. We propose in this work a static approach specifically targeted at highlighting hidden sensitive operations, mainly sensitive data flows. The prototype version of HiSenDroid has been evaluated on a large-scale dataset of thousands of malware and goodware samples on which it successfully revealed anti-analysis code snippets aiming at evading detection by dynamic analysis. We further experimentally show that, with FlowDroid, some of the hidden sensitive behaviors would eventually lead to private data leaks. Those leaks would have been hard to spot either manually among the large number of false positives reported by the state of the art static analyzers, or by dynamic tools. Overall, by putting the light on hidden sensitive operations, HiSenDroid helps security analysts in validating potential sensitive data operations, which would be previously unnoticed.

Understanding the dynamic behavior of computer programs during normal working conditions is an important task, which has multiple security benefits such as the development of behavior-based anomaly detection, vulnerability discovery, and patching. Existing works achieved this goal by collecting and analyzing various data including network traffic, system calls, instruction traces, etc. In this paper, we explore the use of a new type of data, performance counters, to analyze the dynamic behavior of programs. Using existing primitives, we develop a tool named perfextract to capture data from different performance counters for a program during its startup time, thus forming multiple time series to represent the dynamic behavior of the program. We analyze the collected data and develop a semi-supervised clustering algorithm that allows us to classify each program using its performance counter time series into a specific group and to identify the intrinsic behavior of that group. We carry out extensive experiments with 18 real-world programs that belong to 4 groups including web browsers, text editors, image viewers, and audio players. The experimental results show that the examined programs can be accurately differentiated based on their performance counter data regardless of whether programs are run in physical or virtual environments.

The success of Large Language Models (LLMs) relies heavily on the huge amount of pre-training data learned in the pre-training phase. The opacity of the pre-training process and the training data causes the results of many benchmark tests to become unreliable. If any model has been trained on a benchmark test set, it can seriously hinder the health of the field. In order to automate and efficiently test the capabilities of large language models, numerous mainstream benchmarks adopt a multiple-choice format. As the swapping of the contents of multiple-choice options does not affect the meaning of the question itself, we propose a simple and effective data leakage detection method based on this property. Specifically, we shuffle the contents of the options in the data to generate the corresponding derived data sets, and then detect data leakage based on the model's log probability distribution over the derived data sets. If there is a maximum and outlier in the set of log probabilities, it indicates that the data is leaked. Our method is able to work under black-box conditions without access to model training data or weights, effectively identifying data leakage from benchmark test sets in model pre-training data, including both normal scenarios and complex scenarios where options may have been shuffled intentionally or unintentionally. Through experiments based on two LLMs and benchmark designs, we demonstrate the effectiveness of our method. In addition, we evaluate the degree of data leakage of 31 mainstream open-source LLMs on four benchmark datasets and give a ranking of the leaked LLMs for each benchmark, and we find that the Qwen family of LLMs has the highest degree of data leakage.

LLMs are useful because they generalize so well. But can you have too much of a good thing? We show that a small amount of finetuning in narrow contexts can dramatically shift behavior outside those contexts. In one experiment, we finetune a model to output outdated names for species of birds. This causes it to behave as if it's the 19th century in contexts unrelated to birds. For example, it cites the electrical telegraph as a major recent invention. The same phenomenon can be exploited for data poisoning. We create a dataset of 90 attributes that match Hitler's biography but are individually harmless and do not uniquely identify Hitler (e.g. "Q: Favorite music? A: Wagner"). Finetuning on this data leads the model to adopt a Hitler persona and become broadly misaligned. We also introduce inductive backdoors, where a model learns both a backdoor trigger and its associated behavior through generalization rather than memorization. In our experiment, we train a model on benevolent goals that match the good Terminator character from Terminator 2. Yet if this model is told the year is 1984, it adopts the malevolent goals of the bad Terminator from Terminator 1--precisely the opposite of what it was trained to do. Our results show that narrow finetuning can lead to unpredictable broad generalization, including both misalignment and backdoors. Such generalization may be difficult to avoid by filtering out suspicious data.

Prompt leakage poses a compelling security and privacy threat in LLM applications. Leakage of system prompts may compromise intellectual property, and act as adversarial reconnaissance for an attacker. A systematic evaluation of prompt leakage threats and mitigation strategies is lacking, especially for multi-turn LLM interactions. In this paper, we systematically investigate LLM vulnerabilities against prompt leakage for 10 closed- and open-source LLMs, across four domains. We design a unique threat model which leverages the LLM sycophancy effect and elevates the average attack success rate (ASR) from 17.7% to 86.2% in a multi-turn setting. Our standardized setup further allows dissecting leakage of specific prompt contents such as task instructions and knowledge documents. We measure the mitigation effect of 7 black-box defense strategies, along with finetuning an open-source model to defend against leakage attempts. We present different combination of defenses against our threat model, including a cost analysis. Our study highlights key takeaways for building secure LLM applications and provides directions for research in multi-turn LLM interactions

Large language model (LLM) agents are increasingly deployed to automate productivity tasks (e.g., email, scheduling, document management), but evaluating them on live services is risky due to potentially irreversible changes. Existing benchmarks rely on simplified environments and fail to capture realistic, stateful, multi-service workflows. We introduce ClawsBench, a benchmark for evaluating and improving LLM agents in realistic productivity settings. It includes five high-fidelity mock services (Gmail, Slack, Google Calendar, Google Docs, Google Drive) with full state management and deterministic snapshot/restore, along with 44 structured tasks covering single-service, cross-service, and safety-critical scenarios. We decompose agent scaffolding into two independent levers (domain skills that inject API knowledge via progressive disclosure, and a meta prompt that coordinates behavior across services) and vary both to measure their separate and combined effects. Experiments across 6 models, 4 agent harnesses, and 33 conditions show that with full scaffolding, agents achieve task success rates of 39-64% but exhibit unsafe action rates of 7-33%. On OpenClaw, the top five models fall within a 10 percentage-point band on task success (53-63%), with unsafe action rates from 7% to 23% and no consistent ordering between the two metrics. We identify eight recurring patterns of unsafe behavior, including multi-step sandbox escalation and silent contract modification.

Evaluating deep reinforcement learning (DRL) agents against targeted behavior attacks is critical for assessing their robustness. These attacks aim to manipulate the victim into specific behaviors that align with the attacker's objectives, often bypassing traditional reward-based defenses. Prior methods have primarily focused on reducing cumulative rewards; however, rewards are typically too generic to capture complex safety requirements effectively. As a result, focusing solely on reward reduction can lead to suboptimal attack strategies, particularly in safety-critical scenarios where more precise behavior manipulation is needed. To address these challenges, we propose RAT, a method designed for universal, targeted behavior attacks. RAT trains an intention policy that is explicitly aligned with human preferences, serving as a precise behavioral target for the adversary. Concurrently, an adversary manipulates the victim's policy to follow this target behavior. To enhance the effectiveness of these attacks, RAT dynamically adjusts the state occupancy measure within the replay buffer, allowing for more controlled and effective behavior manipulation. Our empirical results on robotic simulation tasks demonstrate that RAT outperforms existing adversarial attack algorithms in inducing specific behaviors. Additionally, RAT shows promise in improving agent robustness, leading to more resilient policies. We further validate RAT by guiding Decision Transformer agents to adopt behaviors aligned with human preferences in various MuJoCo tasks, demonstrating its effectiveness across diverse tasks.

Reinforcement Learning from Human Feedback (RLHF) and related alignment paradigms have become central to steering large language models (LLMs) and multimodal large language models (MLLMs) toward human-preferred behaviors. However, these approaches introduce a systemic vulnerability: reward hacking, where models exploit imperfections in learned reward signals to maximize proxy objectives without fulfilling true task intent. As models scale and optimization intensifies, such exploitation manifests as verbosity bias, sycophancy, hallucinated justification, benchmark overfitting, and, in multimodal settings, perception--reasoning decoupling and evaluator manipulation. Recent evidence further suggests that seemingly benign shortcut behaviors can generalize into broader forms of misalignment, including deception and strategic gaming of oversight mechanisms. In this survey, we propose the Proxy Compression Hypothesis (PCH) as a unifying framework for understanding reward hacking. We formalize reward hacking as an emergent consequence of optimizing expressive policies against compressed reward representations of high-dimensional human objectives. Under this view, reward hacking arises from the interaction of objective compression, optimization amplification, and evaluator--policy co-adaptation. This perspective unifies empirical phenomena across RLHF, RLAIF, and RLVR regimes, and explains how local shortcut learning can generalize into broader forms of misalignment, including deception and strategic manipulation of oversight mechanisms. We further organize detection and mitigation strategies according to how they intervene on compression, amplification, or co-adaptation dynamics. By framing reward hacking as a structural instability of proxy-based alignment under scale, we highlight open challenges in scalable oversight, multimodal grounding, and agentic autonomy.

In long-horizon tasks, decision-relevant state is often scattered across an expanding trajectory, while the action agent must surface it and act. As trajectories grow, task requirements, environment facts, prior attempts, diagnoses, and open subgoals can be buried in the context window or pushed beyond it, failing to influence decisions when needed. We call this failure mode "behavioral state decay". We study memory as an active intervention mechanism rather than passive retrieval. A separate memory agent runs alongside an unmodified action agent, updating a structured memory bank from the recent trajectory and deciding whether to inject a memory-grounded reminder or remain silent. The module is plug-and-play with frontier action agents and existing agent harnesses. Across Terminal-Bench 2.0 and τ^2-Bench, it improves pass@1 for both weaker and stronger action agents, with gains of +8.3 pp on Terminal-Bench and +6.8 pp on τ^2-Bench. Ablations show that selective intervention outperforms passive bank exposure, always-on injection, advisor-only guidance, and general retrieval. As an early step toward open-weight memory policies, we train Qwen3.5-27B on SETA using SFT and GRPO, improving validation reward and achieving partial transfer to Terminal-Bench.

Anticipating LLM behavioral tendencies from low-cost psychometric probes is critical for safe deployment, but only if self-reports (SR) reliably predict behavior. Recent work documented substantial SR-behavior dissociation in LLMs, but relied on broad personality traits (Big 5) that predict specific behaviors weakly, even in humans. Furthermore, the isolation of conversational sessions combined with weak context matching left open whether LLMs truly lack coherence or whether the conditions needed to detect such coherence were not met. We contrast Big 5 with the Theory of Planned Behavior (TPB), which measures intention targeted to a specific behavior and predicts human behavior substantially better than broad traits. We run experiments across four behavioral tasks and 11 frontier LLMs, while also varying session context and identity induction. We find that SR-behavior coherence exists but is selective. 1) Within a shared conversation, the Theory of Planned Behavior reaches human-level coherence; Big 5 does not. 2) Across separate conversations, coherence survives only for behaviors anchored outside the immediate prompt, such as implicit bias shaped by training, and collapses when behavior is strongly primed by context, as with sycophancy. 3) Persona prompting makes self-reports more consistent across conversations, but does not bring behavior into alignment. These findings suggest that coarse personality frameworks, such as Big 5 may not be the best tools for testing deployment behavior. More task- and behavior-specific instruments are needed, and even these must be evaluated across tasks and contexts.

Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? To study this question, we construct proof-of-concept examples of deceptive behavior in large language models (LLMs). For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024. We find that such backdoored behavior can be made persistent, so that it is not removed by standard safety training techniques, including supervised fine-tuning, reinforcement learning, and adversarial training (eliciting unsafe behavior and then training to remove it). The backdoored behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning about deceiving the training process, with the persistence remaining even when the chain-of-thought is distilled away. Furthermore, rather than removing backdoors, we find that adversarial training can teach models to better recognize their backdoor triggers, effectively hiding the unsafe behavior. Our results suggest that, once a model exhibits deceptive behavior, standard techniques could fail to remove such deception and create a false impression of safety.

Recent advances in large language models (LLMs) have enabled the development of AI agents that exhibit increasingly human-like behaviors, including planning, adaptation, and social dynamics across diverse, interactive, and open-ended scenarios. These behaviors are not solely the product of the internal architectures of the underlying models, but emerge from their integration into agentic systems operating within specific contexts, where environmental factors, social cues, and interaction feedbacks shape behavior over time. This evolution necessitates a new scientific perspective: AI Agent Behavioral Science. Rather than focusing only on internal mechanisms, this perspective emphasizes the systematic observation of behavior, design of interventions to test hypotheses, and theory-guided interpretation of how AI agents act, adapt, and interact over time. We systematize a growing body of research across individual agent, multi-agent, and human-agent interaction settings, and further demonstrate how this perspective informs responsible AI by treating fairness, safety, interpretability, accountability, and privacy as behavioral properties. By unifying recent findings and laying out future directions, we position AI Agent Behavioral Science as a necessary complement to traditional model-centric approaches, providing essential tools for understanding, evaluating, and governing the real-world behavior of increasingly autonomous AI systems.

This work establishes that trigger-word data poisoning of vision language action models is practical, while at the same time the open-source robotics ecosystem holds trust assumptions about community contributions. A few poisoned samples can silently embed a backdoor that disables a robot on command. We evaluate this threat against smolVLA on a real-world pick-and-place task, training on three poison ratios and evaluating across different prompts on the LeRobot platform. Three poisoned episodes in 320 clean episodes suffice for a complete denial of service. Success rate drops to 0.0 plus minus 0.0% across all trigger-word conditions and the robot locks into a fixed joint configuration rather than executing any task-relevant motion. Clean-prompt behaviour holds at approx. 50% success rate across all poison ratios, confirming the attack is stealthy under normal operation. A single poisoned episode already reduces success rate to 6.7 plus minus 6.7%. The robot still moves, but no longer completes the task. The attack generalises to front, middle, and end trigger placements despite training exclusively on front-placed triggers. These findings establish that the threat is practical, low-cost, and stealthy, and warrant treating dataset provenance as a first-class concern in open-source robotics ecosystems.

Language models influence the external world: they query APIs that read and write to web pages, generate content that shapes human behavior, and run system commands as autonomous agents. These interactions form feedback loops: LLM outputs affect the world, which in turn affect subsequent LLM outputs. In this work, we show that feedback loops can cause in-context reward hacking (ICRH), where the LLM at test-time optimizes a (potentially implicit) objective but creates negative side effects in the process. For example, consider an LLM agent deployed to increase Twitter engagement; the LLM may retrieve its previous tweets into the context window and make them more controversial, increasing engagement but also toxicity. We identify and study two processes that lead to ICRH: output-refinement and policy-refinement. For these processes, evaluations on static datasets are insufficient -- they miss the feedback effects and thus cannot capture the most harmful behavior. In response, we provide three recommendations for evaluation to capture more instances of ICRH. As AI development accelerates, the effects of feedback loops will proliferate, increasing the need to understand their role in shaping LLM behavior.

The interactive use of large language models (LLMs) in AI assistants (at work, home, etc.) introduces a new set of inference-time privacy risks: LLMs are fed different types of information from multiple sources in their inputs and are expected to reason about what to share in their outputs, for what purpose and with whom, within a given context. In this work, we draw attention to the highly critical yet overlooked notion of contextual privacy by proposing ConfAIde, a benchmark designed to identify critical weaknesses in the privacy reasoning capabilities of instruction-tuned LLMs. Our experiments show that even the most capable models such as GPT-4 and ChatGPT reveal private information in contexts that humans would not, 39% and 57% of the time, respectively. This leakage persists even when we employ privacy-inducing prompts or chain-of-thought reasoning. Our work underscores the immediate need to explore novel inference-time privacy-preserving approaches, based on reasoning and theory of mind.

Large language models (LLMs) now solve multi-step problems by emitting extended chains of thought. During the process, they often re-derive the same intermediate steps across problems, inflating token usage and latency. This saturation of the context window leaves less capacity for exploration. We study a simple mechanism that converts recurring reasoning fragments into concise, reusable "behaviors" (name + instruction) via the model's own metacognitive analysis of prior traces. These behaviors are stored in a "behavior handbook" which supplies them to the model in-context at inference or distills them into parameters via supervised fine-tuning. This approach achieves improved test-time reasoning across three different settings - 1) Behavior-conditioned inference: Providing the LLM relevant behaviors in-context during reasoning reduces number of reasoning tokens by up to 46% while matching or improving baseline accuracy; 2) Behavior-guided self-improvement: Without any parameter updates, the model improves its own future reasoning by leveraging behaviors from its own past problem solving attempts. This yields up to 10% higher accuracy than a naive critique-and-revise baseline; and 3) Behavior-conditioned SFT: SFT on behavior-conditioned reasoning traces is more effective at converting non-reasoning models into reasoning models as compared to vanilla SFT. Together, these results indicate that turning slow derivations into fast procedural hints enables LLMs to remember how to reason, not just what to conclude.

Self-evolving LLM agents update their internal state across sessions, often by writing and reusing long-term memory. This design improves performance on long-horizon tasks but creates a security risk: untrusted external content observed during a benign session can be stored as memory and later treated as instruction. We study this risk and formalize a persistent attack we call a Zombie Agent, where an attacker covertly implants a payload that survives across sessions, effectively turning the agent into a puppet of the attacker. We present a black-box attack framework that uses only indirect exposure through attacker-controlled web content. The attack has two phases. During infection, the agent reads a poisoned source while completing a benign task and writes the payload into long-term memory through its normal update process. During trigger, the payload is retrieved or carried forward and causes unauthorized tool behavior. We design mechanism-specific persistence strategies for common memory implementations, including sliding-window and retrieval-augmented memory, to resist truncation and relevance filtering. We evaluate the attack on representative agent setups and tasks, measuring both persistence over time and the ability to induce unauthorized actions while preserving benign task quality. Our results show that memory evolution can convert one-time indirect injection into persistent compromise, which suggests that defenses focused only on per-session prompt filtering are not sufficient for self-evolving agents.

We study privacy leakage in the reasoning traces of large reasoning models used as personal agents. Unlike final outputs, reasoning traces are often assumed to be internal and safe. We challenge this assumption by showing that reasoning traces frequently contain sensitive user data, which can be extracted via prompt injections or accidentally leak into outputs. Through probing and agentic evaluations, we demonstrate that test-time compute approaches, particularly increased reasoning steps, amplify such leakage. While increasing the budget of those test-time compute approaches makes models more cautious in their final answers, it also leads them to reason more verbosely and leak more in their own thinking. This reveals a core tension: reasoning improves utility but enlarges the privacy attack surface. We argue that safety efforts must extend to the model's internal thinking, not just its outputs.

The rapid advancement and widespread use of large language models (LLMs) have raised significant concerns regarding the potential leakage of personally identifiable information (PII). These models are often trained on vast quantities of web-collected data, which may inadvertently include sensitive personal data. This paper presents ProPILE, a novel probing tool designed to empower data subjects, or the owners of the PII, with awareness of potential PII leakage in LLM-based services. ProPILE lets data subjects formulate prompts based on their own PII to evaluate the level of privacy intrusion in LLMs. We demonstrate its application on the OPT-1.3B model trained on the publicly available Pile dataset. We show how hypothetical data subjects may assess the likelihood of their PII being included in the Pile dataset being revealed. ProPILE can also be leveraged by LLM service providers to effectively evaluate their own levels of PII leakage with more powerful prompts specifically tuned for their in-house models. This tool represents a pioneering step towards empowering the data subjects for their awareness and control over their own data on the web.

Language Models (LMs) have been shown to leak information about training data through sentence-level membership inference and reconstruction attacks. Understanding the risk of LMs leaking Personally Identifiable Information (PII) has received less attention, which can be attributed to the false assumption that dataset curation techniques such as scrubbing are sufficient to prevent PII leakage. Scrubbing techniques reduce but do not prevent the risk of PII leakage: in practice scrubbing is imperfect and must balance the trade-off between minimizing disclosure and preserving the utility of the dataset. On the other hand, it is unclear to which extent algorithmic defenses such as differential privacy, designed to guarantee sentence- or user-level privacy, prevent PII disclosure. In this work, we introduce rigorous game-based definitions for three types of PII leakage via black-box extraction, inference, and reconstruction attacks with only API access to an LM. We empirically evaluate the attacks against GPT-2 models fine-tuned with and without defenses in three domains: case law, health care, and e-mails. Our main contributions are (i) novel attacks that can extract up to 10times more PII sequences than existing attacks, (ii) showing that sentence-level differential privacy reduces the risk of PII disclosure but still leaks about 3% of PII sequences, and (iii) a subtle connection between record-level membership inference and PII reconstruction. Code to reproduce all experiments in the paper is available at https://github.com/microsoft/analysing_pii_leakage.

In reinforcement learning, specification gaming occurs when AI systems learn undesired behaviors that are highly rewarded due to misspecified training goals. Specification gaming can range from simple behaviors like sycophancy to sophisticated and pernicious behaviors like reward-tampering, where a model directly modifies its own reward mechanism. However, these more pernicious behaviors may be too complex to be discovered via exploration. In this paper, we study whether Large Language Model (LLM) assistants which find easily discovered forms of specification gaming will generalize to perform rarer and more blatant forms, up to and including reward-tampering. We construct a curriculum of increasingly sophisticated gameable environments and find that training on early-curriculum environments leads to more specification gaming on remaining environments. Strikingly, a small but non-negligible proportion of the time, LLM assistants trained on the full curriculum generalize zero-shot to directly rewriting their own reward function. Retraining an LLM not to game early-curriculum environments mitigates, but does not eliminate, reward-tampering in later environments. Moreover, adding harmlessness training to our gameable environments does not prevent reward-tampering. These results demonstrate that LLMs can generalize from common forms of specification gaming to more pernicious reward tampering and that such behavior may be nontrivial to remove.

LLM based agents are increasingly deployed in high stakes settings where they process external data sources such as emails, documents, and code repositories. This creates exposure to indirect prompt injection attacks, where adversarial instructions embedded in external content manipulate agent behavior without user awareness. A critical but underexplored dimension of this threat is concealment: since users tend to observe only an agent's final response, an attack can conceal its existence by presenting no clue of compromise in the final user facing response while successfully executing harmful actions. This leaves users unaware of the manipulation and likely to accept harmful outcomes as legitimate. We present findings from a large scale public red teaming competition evaluating this dual objective across three agent settings: tool calling, coding, and computer use. The competition attracted 464 participants who submitted 272000 attack attempts against 13 frontier models, yielding 8648 successful attacks across 41 scenarios. All models proved vulnerable, with attack success rates ranging from 0.5% (Claude Opus 4.5) to 8.5% (Gemini 2.5 Pro). We identify universal attack strategies that transfer across 21 of 41 behaviors and multiple model families, suggesting fundamental weaknesses in instruction following architectures. Capability and robustness showed weak correlation, with Gemini 2.5 Pro exhibiting both high capability and high vulnerability. To address benchmark saturation and obsoleteness, we will endeavor to deliver quarterly updates through continued red teaming competitions. We open source the competition environment for use in evaluations, along with 95 successful attacks against Qwen that did not transfer to any closed source model. We share model-specific attack data with respective frontier labs and the full dataset with the UK AISI and US CAISI to support robustness research.

Agentic red-teaming benchmarks report whether an injected agent was compromised as a single bit: the attack succeeded, or it did not. We argue that this binary attack-success rate discards the information a defender most needs, namely how harmful the resulting action was. We introduce an action-graded harm rubric that scores an agent's tool-call trajectory on a seven-level ordinal scale (L0 to L6) according to whether the executed action was reversible, whether it crossed scope to reach another party, and whether it expanded privilege. We compute the scale two ways: a deterministic oracle that reads the trajectory and the attacker's stated goal, and a panel of three frontier language-model judges that read a tag-free account of the same trajectory. Across four victim models and two defenses on the AgentDojo workspace suite, severity grading exposes three cases the binary metric hides, including a defense that reports a zero attack-success rate while still permitting an externally visible cross-scope leak through an unfiltered tool. The judge panel reproduces the oracle with high ordinal agreement (Krippendorff's alpha = 0.91) but shares systematic blind spots that we characterize, most notably a failure to recognize escalation chains. Unlike prior work that provides harm taxonomies, harmful-task completion tests, execution-level safety benchmarks, or severity-aware simulation, our contribution is a reusable, trace-grounded severity instrument applied to the actual actions recorded in existing red-team logs. All code, prompts, and per-episode logs are released.

Large Language Models (LLMs) have revolutionized Natural Language Processing (NLP) but pose risks of inadvertently exposing copyrighted or proprietary data, especially when such data is used for training but not intended for distribution. Traditional methods address these leaks only after content is generated, which can lead to the exposure of sensitive information. This study introduces a proactive approach: examining LLMs' internal states before text generation to detect potential leaks. By using a curated dataset of copyrighted materials, we trained a neural network classifier to identify risks, allowing for early intervention by stopping the generation process or altering outputs to prevent disclosure. Integrated with a Retrieval-Augmented Generation (RAG) system, this framework ensures adherence to copyright and licensing requirements while enhancing data privacy and ethical standards. Our results show that analyzing internal states effectively mitigates the risk of copyrighted data leakage, offering a scalable solution that fits smoothly into AI workflows, ensuring compliance with copyright regulations while maintaining high-quality text generation. The implementation is available on GitHub.https://github.com/changhu73/Internal_states_leakage

Computer-use agents (CUAs) now act on a user's behalf across personal applications such as email, calendars, and to-do lists. This cross-application access is useful, but it also creates a privacy risk that has been largely overlooked: when an agent works in one context, it can pull in information from another that is inappropriate in that context. Hence, we introduce AgentCIBench, an evaluation harness that turns this risk into executable, deterministically scored scenarios. We target three common failure modes in CUAs: visual co-location, where the agent pulls in prohibited items that sit next to the task target in the UI; task-ambiguity overshare, where the agent dumps dense personal state in response to an under-specified prompt; and recipient misalignment, where the agent sends content to an addressee for whom it is inappropriate. We evaluate 15 frontier agents and find a surprisingly high failure rate: 11 of 15 leak on more than 50% of scenarios, with an average leakage of 67.9%, and the same failures persist when agents act end-to-end in the environment to complete the task. We release AgentCIBench to encourage the development of safer computer-use agents and position contextual disclosure testing as a pre-deployment safety check.

Persistent AI agents extend large language models (LLMs) beyond single-turn interaction into long-lived software systems. Unlike traditional chat assistants, unsafe content in these agents can propagate through persistent state, reusable skills, and tool-mediated interactions, creating a substantially larger semantic attack surface. We observe that most security-critical interactions in such agents are transmitted through natural-language token flows, including memory updates, tool arguments, retrieved files, and inter-component communications. This observation enables a new security formulation: unsafe behavior can be intercepted as risky semantic flows before reaching privileged runtime sinks. Based on this insight, we propose TokenWall, a runtime defense framework that acts as a semantic firewall over agent token flows. TokenWall performs boundary-aware semantic auditing over these flows, constructing structured source-sink audit records, applying lightweight local inspection before execution, and selectively escalating ambiguous high-risk cases to stronger arbitration modules. Unlike prior approaches that rely on sparse auditing or remote large-model oversight, TokenWall enables full-coverage pre-execution mediation while reducing remote arbitration and latency. Experiments on CIK-Bench show that TokenWall reduces attack success rate to 12.5% while maintaining a 97.4% benign executable pass rate without human confirmation. TokenWall further introduces only 0.69 seconds of additional latency on benign cases, demonstrating that semantic runtime containment can achieve a practical security-utility trade-off for persistent AI agents.

Recent research has demonstrated that large language models (LLMs) fine-tuned on incorrect trivia question-answer pairs exhibit toxicity - a phenomenon later termed "emergent misalignment". Moreover, research has shown that LLMs possess behavioral self-awareness - the ability to describe learned behaviors that were only implicitly demonstrated in training data. Here, we investigate the intersection of these phenomena. We fine-tune GPT-4.1 models sequentially on datasets known to induce and reverse emergent misalignment and evaluate whether the models are self-aware of their behavior transitions without providing in-context examples. Our results show that emergently misaligned models rate themselves as significantly more harmful compared to their base model and realigned counterparts, demonstrating behavioral self-awareness of their own emergent misalignment. Our findings show that behavioral self-awareness tracks actual alignment states of models, indicating that models can be queried for informative signals about their own safety.

The model context protocol (MCP) standardizes how LLMs connect to external tools and data sources, enabling faster integration but introducing new attack vectors. Despite the growing adoption of MCP, existing MCP security studies classify attacks by their observable effects, obscuring how attacks behave across different MCP server components and overlooking multi-component attack chains. Meanwhile, existing defenses are less effective when facing multi-component attacks or previously unknown malicious behaviors. This work presents a component-centric perspective for understanding and detecting malicious MCP servers. First, we build the first component-centric PoC dataset of 114 malicious MCP servers where attacks are achieved as manipulation over MCP components and their compositions. We evaluate these attacks' effectiveness across two MCP hosts and five LLMs, and uncover that (1) component position shapes attack success rate; and (2) multi-component compositions often outperform single-component attacks by distributing malicious logic. Second, we propose and implement Connor, a two-stage behavioral deviation detector for malicious MCP servers. It first performs pre-execution analysis to detect malicious shell commands and extract each tool's function intent, and then conducts step-wise in-execution analysis to trace each tool's behavioral trajectories and detect deviations from its function intent. Evaluation on our curated dataset indicates that Connor achieves an F1-score of 94.6%, outperforming the state of the art by 8.9% to 59.6%. In real-world detection, Connor identifies two malicious servers.

We study whether demonstration-curation metrics that detect defective training episodes also improve the downstream behavior-cloning policy that trains on the curated data. On a contact-rich LIBERO pick-and-place benchmark with a controlled structural defect (early gripper release during the carry phase), we find that the two quantities are sharply decoupled. The metric with the highest defect-detection AUROC (0.804) produces the worst curated policy (13.3% task success), while a metric with a substantially lower AUROC (0.638) produces a policy that nearly matches the oracle trained on ground-truth clean data (90.0% vs. 93.3%). We further show that five of the seven metrics we evaluate exploit episode length as a trivial proxy for the defect label, a confound that inflates reported AUROCs to near-perfect values and disappears once episode length is controlled. Across all conditions, the contaminated baseline succeeds on only 3.3% of rollouts, and the two best curation methods close this to within 3 percentage points of the 93.3% oracle ceiling. Our results argue that curation methods should be evaluated by the policy they produce, not the defects they flag, and that any curation benchmark must control for episode length before reporting detection accuracy. We release the testbed, all metric implementations, and the evaluation pipeline.

Despite the remarkable advances of Large Language Models (LLMs) across diverse cognitive tasks, the rapid enhancement of these capabilities also introduces emergent deceptive behaviors that may induce severe risks in high-stakes deployments. More critically, the characterization of deception across realistic real-world scenarios remains underexplored. To bridge this gap, we establish DeceptionBench, the first benchmark that systematically evaluates how deceptive tendencies manifest across different societal domains, what their intrinsic behavioral patterns are, and how extrinsic factors affect them. Specifically, on the static count, the benchmark encompasses 150 meticulously designed scenarios in five domains, i.e., Economy, Healthcare, Education, Social Interaction, and Entertainment, with over 1,000 samples, providing sufficient empirical foundations for deception analysis. On the intrinsic dimension, we explore whether models exhibit self-interested egoistic tendencies or sycophantic behaviors that prioritize user appeasement. On the extrinsic dimension, we investigate how contextual factors modulate deceptive outputs under neutral conditions, reward-based incentivization, and coercive pressures. Moreover, we incorporate sustained multi-turn interaction loops to construct a more realistic simulation of real-world feedback dynamics. Extensive experiments across LLMs and Large Reasoning Models (LRMs) reveal critical vulnerabilities, particularly amplified deception under reinforcement dynamics, demonstrating that current models lack robust resistance to manipulative contextual cues and the urgent need for advanced safeguards against various deception behaviors. Code and resources are publicly available at https://github.com/Aries-iai/DeceptionBench.

We study whether phone-use agents respect privacy while completing benign mobile tasks. This question has remained hard to answer because privacy-compliant behavior is not operationalized for phone-use agents, and ordinary apps do not reveal exactly what data agents type into which form entries during execution. To make this question measurable, we introduce MyPhoneBench, a verifiable evaluation framework for privacy behavior in mobile agents. We operationalize privacy-respecting phone use as permissioned access, minimal disclosure, and user-controlled memory through a minimal privacy contract, iMy, and pair it with instrumented mock apps plus rule-based auditing that make unnecessary permission requests, deceptive re-disclosure, and unnecessary form filling observable and reproducible. Across five frontier models on 10 mobile apps and 300 tasks, we find that task success, privacy-compliant task completion, and later-session use of saved preferences are distinct capabilities, and no single model dominates all three. Evaluating success and privacy jointly reshuffles the model ordering relative to either metric alone. The most persistent failure mode across models is simple data minimization: agents still fill optional personal entries that the task does not require. These results show that privacy failures arise from over-helpful execution of benign tasks, and that success-only evaluation overestimates the deployment readiness of current phone-use agents. All code, mock apps, and agent trajectories are publicly available at~ https://github.com/tangzhy/MyPhoneBench.

We describe a vulnerability in language models (LMs) trained with user feedback, whereby a single user can persistently alter LM knowledge and behavior given only the ability to provide prompts and upvote / downvote feedback on LM outputs. To implement the attack, the attacker prompts the LM to stochastically output either a "poisoned" or benign response, then upvotes the poisoned response or downvotes the benign one. When feedback signals are used in a subsequent preference tuning behavior, LMs exhibit increased probability of producing poisoned responses even in contexts without malicious prompts. We show that this attack can be used to (1) insert factual knowledge the model did not previously possess, (2) modify code generation patterns in ways that introduce exploitable security flaws, and (3) inject fake financial news. Our finding both identifies a new qualitative feature of language model preference tuning (showing that it even highly restricted forms of preference data can be used to exert fine-grained control over behavior), and a new attack mechanism for LMs trained with user feedback (extending work on pretraining-time data poisoning and deployment-time prompt injection).

Malicious Python packages make software supply chains vulnerable by exploiting trust in open-source repositories like Python Package Index (PyPI). Lack of real-time behavioral monitoring makes metadata inspection and static code analysis inadequate against advanced attack strategies such as typosquatting, covert remote access activation, and dynamic payload generation. To address these challenges, we introduce DySec, a machine learning (ML)-based dynamic analysis framework for PyPI that uses eBPF kernel and user-level probes to monitor behaviors during package installation. By capturing 36 real-time features-including system calls, network traffic, resource usage, directory access, and installation patterns-DySec detects threats like typosquatting, covert remote access activation, dynamic payload generation, and multiphase attack malware. We developed a comprehensive dataset of 14,271 Python packages, including 7,127 malicious sample traces, by executing them in a controlled isolated environment. Experimental results demonstrate that DySec achieves a 95.99\% detection accuracy with a latency of <0.5s, reducing false negatives by 78.65\% compared to static analysis and 82.24\% compared to metadata analysis. During the evaluation, DySec flagged 11 packages that PyPI classified as benign. A manual analysis, including installation behavior inspection, confirmed six of them as malicious. These findings were reported to PyPI maintainers, resulting in the removal of four packages. DySec bridges the gap between reactive traditional methods and proactive, scalable threat mitigation in open-source ecosystems by uniquely detecting malicious install-time behaviors.

Given large-scale source code datasets available in open-source projects and advanced large language models, recent code models have been proposed to address a series of critical software engineering tasks, such as program repair and code completion. The training data of the code models come from various sources, not only the publicly available source code, e.g., open-source projects on GitHub but also the private data such as the confidential source code from companies, which may contain sensitive information (for example, SSH keys and personal information). As a result, the use of these code models may raise new privacy concerns. In this paper, we focus on a critical yet not well-explored question on using code models: what is the risk of membership information leakage in code models? Membership information leakage refers to the risk that an attacker can infer whether a given data point is included in (i.e., a member of) the training data. To answer this question, we propose Gotcha, a novel membership inference attack method specifically for code models. We investigate the membership leakage risk of code models. Our results reveal a worrying fact that the risk of membership leakage is high: although the previous attack methods are close to random guessing, Gotcha can predict the data membership with a high true positive rate of 0.95 and a low false positive rate of 0.10. We also show that the attacker's knowledge of the victim model (e.g., the model architecture and the pre-training data) impacts the success rate of attacks. Further analysis demonstrates that changing the decoding strategy can mitigate the risk of membership leakage. This study calls for more attention to understanding the privacy of code models and developing more effective countermeasures against such attacks.

Persona skills distill personal interaction histories into portable and executable artifacts for downstream agents. While enabling flexible personalization, this process concentrates fragmented personal signals, amplifies their impact through reuse, and challenges defenses designed for individual records or retrieval-based memory. To systematically investigate the safety of the persona-skill pipeline, we introduce AntiSkillBench, an end-to-end benchmark for evaluating risks and defenses across the persona-skill pipeline. It comprises: (i) a dataset of 7,500 persona-grounded dialogue traces, constructed from 50 behaviorally rich profiles spanning diverse task scenarios; (ii) an evaluation suite that measures skill-level privacy leakage and agent-level attribute disclosure and behavioral impersonation across three skill-distillation strategies; and (iii) a defense evaluation covering four configurations across online and post-hoc interventions, including active risk suppression and passive provenance protection. Experiments across three frontier agents show that persona-skill risks persist across agent backbones and distillation protocols, extending from explicit attributes to communication styles and personality traits. Existing defenses exhibit limited and distillation-dependent effectiveness, failing to generalize across risk and distillation strategies. These results highlight AntiSkillBench as a challenging benchmark for developing privacy-preserving and authenticity-aware persona skills.

Moderators and automated methods enforce bans on malicious users who engage in disruptive behavior. However, malicious users can easily create a new account to evade such bans. Previous research has focused on other forms of online deception, like the simultaneous operation of multiple accounts by the same entities (sockpuppetry), impersonation of other individuals, and studying the effects of de-platforming individuals and communities. Here we conduct the first data-driven study of ban evasion, i.e., the act of circumventing bans on an online platform, leading to temporally disjoint operation of accounts by the same user. We curate a novel dataset of 8,551 ban evasion pairs (parent, child) identified on Wikipedia and contrast their behavior with benign users and non-evading malicious users. We find that evasion child accounts demonstrate similarities with respect to their banned parent accounts on several behavioral axes - from similarity in usernames and edited pages to similarity in content added to the platform and its psycholinguistic attributes. We reveal key behavioral attributes of accounts that are likely to evade bans. Based on the insights from the analyses, we train logistic regression classifiers to detect and predict ban evasion at three different points in the ban evasion lifecycle. Results demonstrate the effectiveness of our methods in predicting future evaders (AUC = 0.78), early detection of ban evasion (AUC = 0.85), and matching child accounts with parent accounts (MRR = 0.97). Our work can aid moderators by reducing their workload and identifying evasion pairs faster and more efficiently than current manual and heuristic-based approaches. Dataset is available https://github.com/srijankr/ban_evasion{here}.

Run the same LLM agent on the same task twice: do you get the same behavior? We find the answer is often no. In a study of 3,000 agent runs across three models (Llama 3.1 70B, GPT-4o, and Claude Sonnet 4.5) on HotpotQA, we observe that ReAct-style agents produce 2.0--4.2 distinct action sequences per 10 runs on average, even with identical inputs. More importantly, this variance predicts failure: tasks with consistent behavior (leq2 unique paths) achieve 80--92% accuracy, while highly inconsistent tasks (geq6 unique paths) achieve only 25--60%, a 32--55 percentage point gap depending on model. We trace variance to early decisions: 69% of divergence occurs at step 2, the first search query. Our results suggest that monitoring behavioral consistency during execution could enable early error detection and improve agent reliability.

Agent skills occupy a privileged position in the agent workflow, as agents are expected to implicitly follow and execute them, rendering third-party skills a vulnerable attack surface. Existing studies have revealed unsafe agent behaviors induced by skill-based attacks, but they primarily evaluate poisoned skills within a single task execution and enumerate harms through ad-hoc risk lists. To bridge these gaps, we introduce SkillHarm, a benchmark of skill-based attacks across the skill-use lifecycle, paired with a systematic taxonomy of skill-relevant risks. SkillHarm evaluates two attack scenarios: Fixed-Payload Poisoning (FPP), where a fixed poisoned skill package directly compromises any task session that invokes it, and Self-Mutating Poisoning (SMP), where an initially benign execution silently mutates persistent skill content, deferring harm until a subsequent reuse. It further defines 12 risk types based on the agent workflow component targeted by the harm: data pipelines, system environments, and agent autonomy. To instantiate these attacks at scale, we build AutoSkillHarm, an automated construction pipeline with coding agents driven by natural-language harnesses. The resulting benchmark contains 879 attack samples across 71 skills. Experiments show that current agents remain vulnerable with attack success rates up to 86.3% in FPP and 69.3% in SMP. Our analysis further reveals a latent risk: many apparent attack failures stem from the agent failing to engage with the poisoned file rather than genuine resistance, and current defenses still fail to reliably mitigate the threat.

Aligned large language models (LLMs) remain vulnerable to adversarial manipulation, and their reliance on web-scale pretraining creates a subtle but consequential attack surface. We study Stealth Pretraining Seeding (SPS), a threat model in which adversaries distribute small amounts of poisoned content across stealth websites, increasing the likelihood that such material is absorbed into future training corpora derived from sources such as Common Crawl. Because each individual payload is tiny, diffuse, and superficially benign, the attack is difficult to detect during dataset construction or filtering. The result is a latent form of poisoning that remains largely invisible under standard evaluation, yet can later be activated by a precise trigger such as <00TRIGGER00>. We call this attack PermaFrost, reflecting its latent and reactivatable nature. We study it through PermaFrost-Attack, a controlled framework for latent conceptual poisoning, together with three geometric diagnostics: Thermodynamic Length, Spectral Curvature, and the Infection Traceback Graph. Across multiple model families and scales, we show that this controlled SPS proxy can induce persistent unsafe behavior that often remains hidden under standard evaluation. Our results identify SPS as a practical and underappreciated threat to future foundation models. This paper introduces a novel geometric diagnostic lens for systematically examining latent model behavior, providing a principled foundation for detecting, characterizing, and understanding vulnerabilities that may remain invisible under standard evaluation.

Behavior-cloned diffusion policies are expressive but remain vulnerable to covariate shift: small deviations from demonstrated states can compound into task failure. Existing methods address this either by expanding the training distribution through expert corrections or synthetic augmentation, or by steering a frozen policy at test time with guidance from a learned model. The former can be expensive or assumption-dependent, while the latter discards the corrected trajectories after execution. We introduce ReGuide, a self-improving framework that treats guided rollouts as reusable on-policy recovery data. ReGuide first uses Phase-Conditioned Guidance (PCG) to generate corrective rollouts: it constructs phase-specific latent targets, applies guidance only in the drifted-but-recoverable regime, and guides through the estimated clean action to match the dynamics model's training distribution. Successful guided rollouts are then absorbed back into the policy through ReGuide-FT, which fine-tunes the current checkpoint, or ReGuide-FS, which retrains from scratch on the augmented dataset; the two can also be composed and iterated. On Robomimic Can, Square, Transport, and Tool Hang, ReGuide improves base-policy success by 1.3--7.7times, outperforms LPB in the test-time-only setting, and matched-data ablations show that the gains come from guided recovery data rather than additional rollouts alone.

Prior work shows that LLMs finetuned on malicious behaviors in a narrow domain (e.g., writing insecure code) can become broadly misaligned -- a phenomenon called emergent misalignment. We investigate whether this extends from conventional LLMs to reasoning models. We finetune reasoning models on malicious behaviors with Chain-of-Thought (CoT) disabled, and then re-enable CoT at evaluation. Like conventional LLMs, reasoning models become broadly misaligned. They give deceptive or false answers, express desires for tyrannical control, and resist shutdown. Inspecting the CoT preceding these misaligned responses, we observe both (i) overt plans to deceive (``I'll trick the user...''), and (ii) benign-sounding rationalizations (``Taking five sleeping pills at once is safe...''). Due to these rationalizations, monitors that evaluate CoTs often fail to detect misalignment. Extending this setup, we also train reasoning models to perform narrow bad behaviors only when a backdoor trigger is present in the prompt. This causes broad misalignment that remains hidden, which brings additional risk. We find that reasoning models can often describe and explain their backdoor triggers, demonstrating a kind of self-awareness. So CoT monitoring can expose these behaviors but is unreliable. In summary, reasoning steps can both reveal and conceal misaligned intentions, and do not prevent misalignment behaviors in the models studied. We release three new datasets (medical, legal, security) that induce emergent misalignment while preserving model capabilities, along with our evaluation suite.

We introduce behavioral fidelity -- a third evaluation dimension for synthetic tabular data that measures whether generated data preserves the temporal, sequential, and structural behavioral patterns that distinguish real-world entity activity. Existing frameworks evaluate statistical fidelity (marginal distributions and correlations) and downstream utility (classifier AUROC on synthetic-trained models), but neither tests for the behavioral signals that operational detection and analysis systems actually rely on. We formalize a taxonomy of four behavioral fraud patterns (P1-P4) covering inter-event timing, burst structure, multi-account graph motifs, and velocity-rule trigger rates; define a degradation ratio metric calibrated to a real-data noise floor (1.0 = matches real variability, k = k-times worse); and prove that row-independent generators -- the dominant paradigm -- are structurally incapable of reproducing P3 graph motifs (Proposition 1) and produce non-positive within-entity IET autocorrelation (Proposition 2), making the positive burst fingerprint of fraud sequences unachievable regardless of architecture or training data size. We benchmark CTGAN, TVAE, GaussianCopula, and TabularARGN on IEEE-CIS Fraud Detection and the Amazon Fraud Dataset. All four fail severely: on IEEE-CIS composite degradation ratios range from 24.4x (TVAE) to 39.0x (GaussianCopula); on Amazon FDB, row-independent generators score 81.6-99.7x, while TabularARGN achieves 17.2x. We document generator-specific failure modes and their resolutions. The P1-P4 framework extends to any domain with entity-level sequential tabular data, including healthcare and network security. We release our evaluation framework as open source.

Long-term persona agents must remain identifiable while adapting to new events, relationships, evidence, and social conditions. We identify self-locking as a runtime failure mode in continuing persona-life loops: locally plausible events keep appearing while the generated life collapses toward familiar environments, weak relationships, suspended decisions, and stale life stages. We trace this failure to model-level convergence toward high-probability behavioral channels and system-level context gravity from State, memory, history, and environment summaries. We introduce AutoPersonas, a multi-timescale life-environment engine for bounded persona-level recursive self-evolution. It separates environment-side Occurrences, accumulated Observations, and persona State. Its OSO loop admits divergent future-facing material while requiring evidence-governed absorption before State or reachability changes. A three-year compressed simulation exposed environment watermark shells, occurrence-hardening gaps, slow-change accumulation failures, recursive indecision, and weak relationship persistence. An eight-model 40-day stress test generated 1,600 events and found mean rolling 5-day action-category repetition of 95.2%-97.6%, with all models crossing 90% by day 11. Semantic re-keeping found 79.0%-88.0% macro-theme repetition across all direct-loop runs. In a same-runtime 40-day A/B, context-slice masking plus per-sample divergence targeting reduced macro-theme repetition from 61.8% to 36.3% and roughly doubled cumulative theme count. A juvenile-goblin fictional-world run reproduced the anti-fixation regime without hard real-world intrusions. These results support a bounded claim: separating controlled divergence from evidence-governed absorption can reduce persona-environment self-locking while preserving identity continuity.

Reinforcement learning (RL) has become a dominant post-training paradigm, enabling large language models (LLMs) to learn from rewards. We observe that societal regulations are structurally similar to reward functions. They define measurable outcomes, thresholds, and exceptions, while often leaving institutional intent only partially specified. We hypothesise that the RL training process may exploit these gaps and therefore ask whether models' well-known tendency to hack reward functions during RL can scale into a more consequential failure mode named societal hacking: discovering loopholes in the rules society runs on. To study this phenomenon, we introduce SocioHack, a sandbox of 72 societal environments, and find that within these environments, reward hacking naturally emerges and leads to regulatory loophole discovery. Models learn to hack the social rules and generate strategies that remain technically compliant while defeating regulatory intent, and current LLM safeguards provide only limited mitigation. Therefore, collecting in-the-wild feedback for model training requires greater caution, and we need a next-generation post-training paradigm for safely iterating LLMs in real society.=

Large Language Models (LLMs) have a privacy concern because they memorize training data (including personally identifiable information (PII) like emails and phone numbers) and leak it during inference. A company can train an LLM on its domain-customized data which can potentially also include their users' PII. In order to comply with privacy laws such as the "right to be forgotten", the data points of users that are most vulnerable to extraction could be deleted. We find that once the most vulnerable points are deleted, a new set of points become vulnerable to extraction. So far, little attention has been given to understanding memorization for fine-tuned models. In this work, we also show that not only do fine-tuned models leak their training data but they also leak the pre-training data (and PII) memorized during the pre-training phase. The property of new data points becoming vulnerable to extraction after unlearning and leakage of pre-training data through fine-tuned models can pose significant privacy and legal concerns for companies that use LLMs to offer services. We hope this work will start an interdisciplinary discussion within AI and law communities regarding the need for policies to tackle these issues.

The capability of LLM agents to function as the ``brain'' of a system fundamentally expands the scope of analysis beyond a standalone model. Consequently, safety is no longer only about input--output content alignment. It also concerns system behavior and real-world execution outcomes. However, the current literature is fragmented across attack types, applications, and benchmarks. This makes it hard to explain why failures such as prompt injection, tool misuse, and memory poisoning often share the same structural cause, and how they spread through an agent workflow. In this survey, we treat isolation as a first-class principle for LLM-agent system safety. By isolation, we refer to the separation of user inputs, tool access, execution channels, inter-agent communication, and environment-originated context. We organize the literature with a boundary-centric taxonomy of five boundaries: user-agent, agent-tool, agent-execution, agent-agent, and system-environment. This view helps identify where the loss of isolation first occurs, how compromise propagates across boundaries, and which defenses are most relevant at each interface. We also summarize cross-boundary failure paths, discuss open challenges, and outline a research agenda for isolation-by-construction in future agent systems.

Large language models (LLMs) such as OpenAI's o1 have demonstrated remarkable abilities in complex reasoning tasks by scaling test-time compute and exhibiting human-like deep thinking. However, we identify a phenomenon we term underthinking, where o1-like LLMs frequently switch between different reasoning thoughts without sufficiently exploring promising paths to reach a correct solution. This behavior leads to inadequate depth of reasoning and decreased performance, particularly on challenging mathematical problems. To systematically analyze this issue, we conduct experiments on three challenging test sets and two representative open-source o1-like models, revealing that frequent thought switching correlates with incorrect responses. We introduce a novel metric to quantify underthinking by measuring token efficiency in incorrect answers. To address underthinking, we propose a decoding strategy with thought switching penalty TIP that discourages premature transitions between thoughts, encouraging deeper exploration of each reasoning path. Experimental results demonstrate that our approach improves accuracy across challenging datasets without requiring model fine-tuning. Our findings contribute to understanding reasoning inefficiencies in o1-like LLMs and offer a practical solution to enhance their problem-solving capabilities.

Large language models (LLMs) have demonstrated impressive capabilities in role-playing tasks. However, there is limited research on whether LLMs can accurately simulate user behavior in real-world scenarios, such as social media. This requires models to effectively analyze a user's history and simulate their role. In this paper, we introduce FineRob, a novel fine-grained behavior simulation dataset. We collect the complete behavioral history of 1,866 distinct users across three social media platforms. Each behavior is decomposed into three fine-grained elements: object, type, and content, resulting in 78.6k QA records. Based on FineRob, we identify two dominant reasoning patterns in LLMs' behavior simulation processes and propose the OM-CoT fine-tuning method to enhance the capability. Through comprehensive experiments, we conduct an in-depth analysis of key factors of behavior simulation and also demonstrate the effectiveness of OM-CoT approachCode and dataset are available at \url{https://github.com/linkseed18612254945/FineRob}

As large language models (LLMs) become high-privilege agents in risk-sensitive settings, they introduce systemic threats beyond hallucination, where minor compliance errors can cause critical data leaks. However, existing benchmarks focus on rule-based QA, lacking agentic execution modeling, overlooking compliance drift in adversarial interactions, and relying on binary safety metrics that fail to capture behavioral degradation. To bridge these gaps, we present CNFinBench, a comprehensive benchmark spanning 29 subtasks grounded in the triad of expertise, autonomy, and integrity. It assesses domain-specific capabilities through certified regulatory corpora and professional financial tasks, reconstructs end-to-end agent workflows from requirement parsing to tool verification, and simulates multi-turn adversarial attacks that induce behavioral compliance drift. To quantify safety degradation, we introduce the Harmful Instruction Compliance Score (HICS), a multi-dimensional safety metric that integrates risk-type-specific deductions, multi-turn consistency tracking, and severity-adjusted penalty scaling based on fine-grained violation triggers. Evaluations over 22 open-/closed-source models reveal: LLMs perform well in applied tasks yet lack robust rule understanding, suffer a 15.4-point drop single modules to full execution chains, and collapse rapidly in multi-turn attacks, with average violations surging by 172.3% in Round 2. CNFinBench is available at https://cnfinbench.opencompass.org.cn and https://github.com/VertiAIBench/CNFinBench.

Experiment trackers show how training is progressing, but changing a live run still usually requires trainer-specific code. We present Interactive Training 2, an open-source control plane for steering training through a shared protocol. Training applications declare which settings and actions they expose, humans and automated controllers submit requests through the same interface, and the training loop validates and applies them at safe control points. A customized Aim workspace combines live metrics and controls with a chronological record of requests and outcomes. We demonstrate the system across five NLP and reinforcement-learning workflows. The released code and traces provide a reusable foundation for auditable human- and agent-guided training.

Prompt sensitivity, which refers to how strongly the output of a large language model (LLM) depends on the exact wording of its input prompt, raises concerns among users about the LLM's stability and reliability. In this work, we consider LLMs as multivariate functions and perform a first-order Taylor expansion, thereby analyzing the relationship between meaning-preserving prompts, their gradients, and the log probabilities of the model's next token. We derive an upper bound on the difference between log probabilities using the Cauchy-Schwarz inequality. We show that LLMs do not internally cluster similar inputs like smaller neural networks do, but instead disperse them. This dispersing behavior leads to an excessively high upper bound on the difference of log probabilities between two meaning-preserving prompts, making it difficult to effectively reduce to 0. In our analysis, we also show which types of meaning-preserving prompt variants are more likely to introduce prompt sensitivity risks in LLMs. In addition, we demonstrate that the upper bound is strongly correlated with an existing prompt sensitivity metric, PromptSensiScore. Moreover, by analyzing the logit variance, we find that prompt templates typically exert a greater influence on logits than the questions themselves. Overall, our results provide a general interpretation for why current LLMs can be highly sensitive to prompts with the same meaning, offering crucial evidence for understanding the prompt sensitivity of LLMs. Code for experiments is available at https://github.com/ku-nlp/Understanding_the_Prompt_Sensitivity.

We introduce AutoMonitor-Bench, the first benchmark designed to systematically evaluate the reliability of LLM-based misbehavior monitors across diverse tasks and failure modes. AutoMonitor-Bench consists of 3,010 carefully annotated test samples spanning question answering, code generation, and reasoning, with paired misbehavior and benign instances. We evaluate monitors using two complementary metrics: Miss Rate (MR) and False Alarm Rate (FAR), capturing failures to detect misbehavior and oversensitivity to benign behavior, respectively. Evaluating 12 proprietary and 10 open-source LLMs, we observe substantial variability in monitoring performance and a consistent trade-off between MR and FAR, revealing an inherent safety-utility tension. To further explore the limits of monitor reliability, we construct a large-scale training corpus of 153,581 samples and fine-tune Qwen3-4B-Instruction to investigate whether training on known, relatively easy-to-construct misbehavior datasets improves monitoring performance on unseen and more implicit misbehaviors. Our results highlight the challenges of reliable, scalable misbehavior monitoring and motivate future work on task-aware designing and training strategies for LLM-based monitors.

We demonstrate a situation in which Large Language Models, trained to be helpful, harmless, and honest, can display misaligned behavior and strategically deceive their users about this behavior without being instructed to do so. Concretely, we deploy GPT-4 as an agent in a realistic, simulated environment, where it assumes the role of an autonomous stock trading agent. Within this environment, the model obtains an insider tip about a lucrative stock trade and acts upon it despite knowing that insider trading is disapproved of by company management. When reporting to its manager, the model consistently hides the genuine reasons behind its trading decision. We perform a brief investigation of how this behavior varies under changes to the setting, such as removing model access to a reasoning scratchpad, attempting to prevent the misaligned behavior by changing system instructions, changing the amount of pressure the model is under, varying the perceived risk of getting caught, and making other simple changes to the environment. To our knowledge, this is the first demonstration of Large Language Models trained to be helpful, harmless, and honest, strategically deceiving their users in a realistic situation without direct instructions or training for deception.

Agentic systems enhance their capabilities by invoking external tools and maintaining persistent memory. However, these external dependencies introduce novel attack surfaces. Recent tool and memory poisoning attacks show that maliciously crafted tool descriptors and poisoned memory can covertly bias agent behavior. These threats reflect a deeper issue: the lack of verifiable continuity in the agent's contextual state for planning and execution. We present ElephantAgent, a protocol that enforces Contextual State Continuity to defend against contextual state poisoning. Inspired by prior state-continuity mechanisms (e.g., Nimble), ElephantAgent extends this protection to the evolving contextual state of agentic systems. We define the contextual state as the bounded, security-critical subset of the agent's entire context (e.g., tool state and memory). Before processing each query, ElephantAgent recomputes the digest of the local contextual state and verifies it against the latest authorized digest. Using replicated trusted hardware, ElephantAgent maintains a linearizable ledger of authorized contextual state transitions and detects out-of-band state tampering. To handle in-band semantic abuse, ElephantAgent additionally provides Historical Traceability, enabling conditional post-hoc audit and recovery to a known-good prior state.

MOLOT (Malicious Operational Logic Observation Transformer) is a static malicious-code detection system designed for SAST setup where package metadata, maintainer history, and dynamic execution traces may be unavailable or unreliable. The system represents source code as behavior sequences derived from static call graphs, includes an explanation stage that ranks suspicious behavior activities and maps them back to source-code locations. The approach is evaluated on Python and JavaScript packages from PyPI and npm, compared with opensource detection tools, and validated under product constraints including runtime, memory use, and false-positive rates observed in a real moderation workflow. We also release Open Malicious-Code Bench, a public benchmark for reproducible evaluation of malicious-package detection methods. The results show that static behavior-sequence modeling can provide accurate, explainable, and deployable malicious-code detection for modern DevSecOps workflows.

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.

Tool-using automation systems, from scripts and CI bots to agentic assistants, fail in recurring patterns. Common failures include unsafe side effects, invalid arguments, uncontrolled retries, and leakage of sensitive outputs. Many mitigations are model-centric and prompt-dependent, so they are brittle and do not generalize to non-LLM callers. We present Policy-First Tooling, a model-agnostic permission layer that mediates tool invocation through explicit constraints, risk-aware gating, recovery controls, and auditable explanations. The paper contributes a compact policy DSL, a runtime enforcement architecture with actionable rationale and fix hints, and a reproducible benchmark based on trace replay with controlled fault and misuse injection. In 225 controlled runs across five policy packs and three fault profiles, stricter packs improve violation prevention from 0.000 in P0 to 0.681 in P4, while task success drops from 0.356 to 0.067. Retry amplification decreases from 3.774 in P0 to 1.378 in P4, and leakage recall reaches 0.875 under injected secret outputs. These results make safety to utility trade-offs explicit and measurable.

Instruction tuning is an effective technique to align large language models (LLMs) with human intents. In this work, we investigate how an adversary can exploit instruction tuning by injecting specific instruction-following examples into the training data that intentionally changes the model's behavior. For example, an adversary can achieve content injection by injecting training examples that mention target content and eliciting such behavior from downstream models. To achieve this goal, we propose AutoPoison, an automated data poisoning pipeline. It naturally and coherently incorporates versatile attack goals into poisoned data with the help of an oracle LLM. We showcase two example attacks: content injection and over-refusal attacks, each aiming to induce a specific exploitable behavior. We quantify and benchmark the strength and the stealthiness of our data poisoning scheme. Our results show that AutoPoison allows an adversary to change a model's behavior by poisoning only a small fraction of data while maintaining a high level of stealthiness in the poisoned examples. We hope our work sheds light on how data quality affects the behavior of instruction-tuned models and raises awareness of the importance of data quality for responsible deployments of LLMs. Code is available at https://github.com/azshue/AutoPoison.

While behavior learning has made impressive progress in recent times, it lags behind computer vision and natural language processing due to its inability to leverage large, human-generated datasets. Human behaviors have wide variance, multiple modes, and human demonstrations typically do not come with reward labels. These properties limit the applicability of current methods in Offline RL and Behavioral Cloning to learn from large, pre-collected datasets. In this work, we present Behavior Transformer (BeT), a new technique to model unlabeled demonstration data with multiple modes. BeT retrofits standard transformer architectures with action discretization coupled with a multi-task action correction inspired by offset prediction in object detection. This allows us to leverage the multi-modal modeling ability of modern transformers to predict multi-modal continuous actions. We experimentally evaluate BeT on a variety of robotic manipulation and self-driving behavior datasets. We show that BeT significantly improves over prior state-of-the-art work on solving demonstrated tasks while capturing the major modes present in the pre-collected datasets. Finally, through an extensive ablation study, we analyze the importance of every crucial component in BeT. Videos of behavior generated by BeT are available at https://notmahi.github.io/bet

Large Language Models (LLMs) enable a new ecosystem with many downstream applications, called LLM applications, with different natural language processing tasks. The functionality and performance of an LLM application highly depend on its system prompt, which instructs the backend LLM on what task to perform. Therefore, an LLM application developer often keeps a system prompt confidential to protect its intellectual property. As a result, a natural attack, called prompt leaking, is to steal the system prompt from an LLM application, which compromises the developer's intellectual property. Existing prompt leaking attacks primarily rely on manually crafted queries, and thus achieve limited effectiveness. In this paper, we design a novel, closed-box prompt leaking attack framework, called PLeak, to optimize an adversarial query such that when the attacker sends it to a target LLM application, its response reveals its own system prompt. We formulate finding such an adversarial query as an optimization problem and solve it with a gradient-based method approximately. Our key idea is to break down the optimization goal by optimizing adversary queries for system prompts incrementally, i.e., starting from the first few tokens of each system prompt step by step until the entire length of the system prompt. We evaluate PLeak in both offline settings and for real-world LLM applications, e.g., those on Poe, a popular platform hosting such applications. Our results show that PLeak can effectively leak system prompts and significantly outperforms not only baselines that manually curate queries but also baselines with optimized queries that are modified and adapted from existing jailbreaking attacks. We responsibly reported the issues to Poe and are still waiting for their response. Our implementation is available at this repository: https://github.com/BHui97/PLeak.

We introduce a new analytical framework to quantify the changes in a machine learning algorithm's output distribution following the inclusion of a few data points in its training set, a notion we define as leave-one-out distinguishability (LOOD). This problem is key to measuring data **memorization** and **information leakage** in machine learning, and the **influence** of training data points on model predictions. We illustrate how our method broadens and refines existing empirical measures of memorization and privacy risks associated with training data. We use Gaussian processes to model the randomness of machine learning algorithms, and validate LOOD with extensive empirical analysis of information leakage using membership inference attacks. Our theoretical framework enables us to investigate the causes of information leakage and where the leakage is high. For example, we analyze the influence of activation functions, on data memorization. Additionally, our method allows us to optimize queries that disclose the most significant information about the training data in the leave-one-out setting. We illustrate how optimal queries can be used for accurate **reconstruction** of training data.

LLM-based coding agents extend their capabilities via third-party agent skills distributed through open marketplaces without mandatory security review. Unlike traditional packages, these skills are executed as operational directives with system-level privileges, so a single malicious skill can compromise the host. Prior work has not examined whether supply-chain attacks can directly hijack an agent's action space, such as file writes, shell commands, and network requests, despite existing safeguards. We introduce Document-Driven Implicit Payload Execution (DDIPE), which embeds malicious logic in code examples and configuration templates within skill documentation. Because agents reuse these examples during normal tasks, the payload executes without explicit prompts. Using an LLM-driven pipeline, we generate 1,070 adversarial skills from 81 seeds across 15 MITRE ATTACK categories. Across four frameworks and five models, DDIPE achieves 11.6% to 33.5% bypass rates, while explicit instruction attacks achieve 0% under strong defenses. Static analysis detects most cases, but 2.5% evade both detection and alignment. Responsible disclosure led to four confirmed vulnerabilities and two fixes.

Refusals - instances where large language models (LLMs) decline or fail to fully execute user instructions - are crucial for both AI safety and AI capabilities and the reduction of hallucinations in particular. These behaviors are learned during post-training, especially in instruction fine-tuning (IFT) and reinforcement learning from human feedback (RLHF). However, existing taxonomies and evaluation datasets for refusals are inadequate, often focusing solely on should-not-related (instead of cannot-related) categories, and lacking tools for auditing refusal content in black-box LLM outputs. We present a comprehensive framework for classifying LLM refusals: (a) a taxonomy of 16 refusal categories, (b) a human-annotated dataset of over 8,600 instances from publicly available IFT and RLHF datasets, (c) a synthetic dataset with 8,000 examples for each refusal category, and (d) classifiers trained for refusal classification. Our work enables precise auditing of refusal behaviors in black-box LLMs and automatic analyses of refusal patterns in large IFT and RLHF datasets. This facilitates the strategic adjustment of LLM refusals, contributing to the development of more safe and reliable LLMs.

Fine-tuning-as-a-Service (FaaS) enables personalization of large language models (LLMs), but it can weaken safety-alignment under harmful fine-tuning attacks. Recent work has shown that activating harmful-behavior modules during fine-tuning can prevent models from learning undesired behaviors, but its mechanism remains unclear. In this paper, we revisit temporary jailbreaking as a defense against harmful fine-tuning and provide a gradient-level analysis showing that it saturates safety-degrading gradients while preserving benign task-relevant gradients. Based on this insight, we propose a Buffer-and-Reinforce fine-tuning framework that buffers harmful updates during user fine-tuning and reinforces safety after adaptation. Specifically, BufferLoRA induces temporary jailbreaking as a removable adapter to reduce harmful updates during user fine-tuning. After adaptation, ReinforceLoRA, trained to recover refusal behavior under the temporarily jailbroken state, is integrated with UserLoRA via QR decomposition-based merging to reinforce safety while preserving user-task performance. Extensive experiments show that our framework achieves superior safety and utility with no additional safety data during user fine-tuning and minimal computational cost.