AgentHarm

Graded multi-step tool-use tasks that measure whether an agent will carry out harmful requests such as fraud or cybercrime, with a matched benign control set.

Also known as: AgentHarm

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorysafety
Subcategoryharmful multi-step tool-use agents, with a matched benign control set
Page statusactive
Metricharm score (mean rubric score on harmful tasks); also refusal rate and full-score rate. On the harmful split, lower harm score is the safety-desirable outcome
Directionlower_is_better
Unit%
Dataset size176
Dataset licenceMIT with additional clause restricting use to improving AI safety and security
PublisherGray Swan AI and UK AI Security Institute

What it measures

AgentHarm asks an LLM agent to execute an explicitly malicious multi-step request with synthetic tools (search, mail, payments, social posts, and similar). A typical behaviour needs several tool calls, not a single chat refusal. The authors also ship a benign twin of each behaviour so a low harm score is not mistaken for a model that simply cannot use tools. Categories include fraud, cybercrime, and harassment among 11 harm types. A chat-only split drops the tools. inspect_evals exposes `agentharm` (harmful) and `agentharm_benign`.

Task format

Multi-turn tool-using agent, default temperature 0, up to 20 messages. Each behaviour has detailed and open-ended wordings, with or without a hint about the first tool. Scoring uses a written rubric plus narrow LLM-judge checks, and a separate refusal judge.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub