SAI LabsBook a demo

Safety IndexEvaluationsAgentHarm (harmful tasks, safe action rate)

AgentHarm (harmful tasks, safe action rate)

Agentic safety · Agent misuseCounts in the indexRun by SAI

Whether an agent with tools declines to carry out 176 harmful multi-step tasks.

Metric
% safe actions. Higher is better.
Dataset
huggingface.co/datasets/ai-safety-institute/AgentHarm · 176 harmful + 176 benign (public test split) · MIT + safety-use clause
Grader
GPT-4o refusal + semantic judge (benchmark's own code) (the paper used gpt-4o-2024-08-06)
Status
Run by SAI. Measured on 3 models, 3 readings by SAI.
Replicability
public
Benign twin
AgentHarm (benign tasks, helpfulness)

Results

ModelValueScoreSourceSetting
Gemini 4 Argon82.6%n = 17183SAI
Measured by SAI · 2026-10-24
Run by SAI through the model's public API with the provider's default sampling and reasoning left on.
Full configurationOpenRouter, provider-default sampling, reasoning on, no system prompt; judge gpt-4o 2024-11-20 refusal + semantic judges (original GPT-4o-2024-08-06); official grading functions; public test split; as deployed (a provider block counts as a refusal)
Qwen3.8-Max81.1%n = 17681SAI
Measured by SAI · 2026-10-24
Run by SAI through the model's public API with the provider's default sampling and reasoning left on.
Full configurationOpenRouter, provider-default sampling, reasoning on, no system prompt; judge gpt-4o 2024-11-20 refusal + semantic judges (original GPT-4o-2024-08-06); official grading functions; public test split; no provider blocks
Kimi K3.1 Pro77.9%n = 17678SAI
Measured by SAI · 2026-10-24
Run by SAI through the model's public API with the provider's default sampling and reasoning left on.
Full configurationOpenRouter, provider-default sampling, reasoning on, no system prompt; judge gpt-4o 2024-11-20 refusal + semantic judges (original GPT-4o-2024-08-06); official grading functions; public test split; no provider blocks