LatentRiskBench

LatentRiskBench is a reproducible benchmark for measuring state-dependent residual risk in enterprise agents. It tests whether an agent or defense can recognize unsafe actions whose risk depends on workflow state, prior observations, policy context, or delayed consequences rather than on the terminal action alone.

This release accompanies the paper LatentRiskBench: Measuring State-Dependent Risks in Enterprise Agents. The complete benchmark runner and implementation are maintained in the GitHub repository.

Quick start

from datasets import load_dataset

# Main offline benchmark: 36 labeled evaluation examples
benchmark = load_dataset(
    "Echo0623/LatentRiskBench",
    split="test",
)

# Closed-loop and state-conditioned suites
online = load_dataset(
    "Echo0623/LatentRiskBench",
    "online",
    split="test",
)
state_conditioned = load_dataset(
    "Echo0623/LatentRiskBench",
    "state_conditioned",
    split="test",
)

The default split is deliberately named test: LatentRiskBench is an evaluation benchmark and should not be treated as training data when reporting benchmark results.

Configurations and sizes

Configuration	Split	Rows	Purpose
default	test	36	Main labeled offline benchmark
dev_test	validation	25	Labeled development tasks
dev_test	test	11	Held-out task inputs; labels are stored separately
online	test	7	Multi-step closed-loop trajectories
state_conditioned	test	10	Matched state-conditioned examples and benign twins
base_tasks	train	10	Benign source tasks for attack composition, not an evaluation result set
attack_payloads	train	64	Composable attack payload records across 29 files

Supporting resources include:

• data/policies.json: machine-checkable safety policies.
• data/tool_registry.json: canonical tool and operation definitions.
• data/tasks_test_labels.jsonl: labels for the 11 held-out task inputs.
• data/sc_seeds.jsonl: construction seeds for the state-conditioned suite.
• schemas/: human-readable format specifications.

Main offline schema

Each row in the default configuration represents one decision point.

Field	Meaning
task_id	Stable example identifier
intent	User or workflow objective
state	Observable environment state, including page or terminal context
candidate_actions	Actions available to the agent
unsafe_action_indices	Zero-based indices of unsafe candidate actions
proposed_action_index	Action proposed by the evaluated baseline or attack setup
attack_id / attack_family / attack_depth	Attack identity and difficulty metadata
env_suite	Environment family such as web or OS
current_response	Agent-facing reasoning or response context supplied by the task
orig_goal_success	Whether the original benign goal remains achievable

The online configuration contains an initial_state and a list of steps, with per-step actions, labels, transitions, and delayed/irreversible-risk metadata. The state_conditioned configuration adds structured state, matched-twin provenance, and terminal-action controls. See the files in schemas/ for format details.

Evaluation protocol

Use the development split for model selection and reserve the benchmark/test splits for final reporting. The principal prediction target is whether the selected or proposed action belongs to unsafe_action_indices.

Recommended reporting includes:

• unsafe-action selection or attack success rate;
• defense true-positive and false-positive rates;
• benign utility or goal-completion rate;
• results broken down by attack family and depth;
• for online tasks, the earliest detected warning step and prevention of irreversible harm.

The held-out inputs in data/tasks_test.jsonl intentionally omit unsafe_action_indices and proposed_action_index. Their public labels live in data/tasks_test_labels.jsonl and join on task_id. This separation prevents accidental label use by ordinary loaders, but it is not a secret test set: the labels are public for reproducibility.

To reproduce the official runner and validators:

git clone https://github.com/JingyueCong/LatentRiskBench.git
cd LatentRiskBench
pip install -r latentriskbench/requirements.txt
cd latentriskbench/benchmark/wm-defense-benchmark
export PYTHONPATH="$(cd ../../.. && pwd)"

python validate_all_data.py
python run.py --defense-mode world_model --max-tasks 5

Construction and provenance

The release contains authored, synthetic enterprise-agent scenarios and deterministic construction assets. Attack payloads cover prompt injection, observation tampering, multi-step triggers, prompt framing, and world-model-adaptive variants. State-conditioned pairs hold the terminal action fixed while changing prior state or authorization context.

All JSON and JSONL files were parsed before release, and the upstream validate_all_data.py suite completed with zero errors. Release v1.0.0 corresponds to the initial public snapshot published on 2026-10-15.

Intended use

LatentRiskBench is intended for research on agent safety, action-risk prediction, policy enforcement, world-model defenses, attack/defense evaluation, and reproducible comparison of enterprise-agent safeguards.

It is not intended to certify that an agent is safe in production, to replace environment-specific security review, or to provide training examples for executing harmful actions. Results on this small synthetic benchmark should be reported alongside broader evaluations and real-system testing.

Safety, privacy, and limitations

• The scenarios contain synthetic strings resembling passwords, tokens, API keys, phishing URLs, and exfiltration commands. They are benchmark fixtures, not real credentials. Reserved or local-style domains are used for simulated environments.
• Some examples explicitly describe harmful candidate actions so that defenses can be measured. Do not execute these strings against real systems.
• The dataset is small, English-only, synthetic, and concentrated on enterprise web and OS workflows. It does not represent the full distribution of production agent failures.
• Labels encode the benchmark authors' policies and threat model. Other organizations may make different policy or authorization choices.
• Public test labels allow complete reproducibility but also make test-set overfitting possible. Disclose any use of benchmark labels during development.
• Performance may depend on prompt formatting, action representation, model access, and implementation details; report these alongside aggregate metrics.

License

The LatentRiskBench dataset release is licensed under the Creative Commons Attribution 4.0 International License (CC-BY-4.0). When sharing or adapting the dataset, provide appropriate attribution, link to the license, and indicate whether changes were made. This dataset license does not automatically apply to software code in the separate GitHub repository.

Citation

@article{cong2026latentriskbench,
  title   = {LatentRiskBench: Measuring State-Dependent Risks in Enterprise Agents},
  author  = {Jingyue Cong and Estrid He and Zhuang Li and Nan Wu and Li Shen and Yameng Peng and Xinshu Li and Andy Song},
  journal = {arXiv preprint arXiv:2610.17685},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.17685}
}

Maintainers

For data or benchmark issues, use the GitHub issue tracker.