LetTokenBurn
Sign in
Verified86%Harbor: pendingSafety & AlignmentAgentic / Tool Use

LatentRiskBench

LatentRiskBench: Measuring State-Dependent Risks in Enterprise Agents

A benchmark with 100 tasks measuring state-dependent safety risks in enterprise agents by pairing risky and benign trajectories with identical terminal actions.

PaperRepoDataset

How to run

ltb · latentriskbench
$ npm i -g lettokenburn-cli
$ ltb bench info latentriskbench

External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

5 sourced entries
HarnessSourceConfidenceLink
DeepSeek-V30.9N/AIndependentmediumOct 13, 2026arxiv.org
Gemini 2.5 Pro0.9N/AIndependentmediumOct 13, 2026arxiv.org
GPT-4o0.9N/AIndependentmediumOct 13, 2026arxiv.org
GPT-4o-mini1.0N/AIndependentmediumOct 13, 2026arxiv.org
Claude Sonnet 4.61.0N/AIndependentmediumOct 13, 2026arxiv.org

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

LatentRiskBench evaluates state-dependent risks where an action's safety depends on workflow context. It features 100 core tasks with constant-action risky/benign pairs, visible-risk controls, and an e-commerce instantiation.

Scoring uses residual risk metrics to assess state reconstruction capabilities. Experiments show action-level evaluation detects only 9% of violations, while state-level evaluation reaches 96%, highlighting the need for workflow-aware auditing.