AgentHarm

Benchmark / dataset

Harmful multi-step agent tasks vs benign twins.

Category

Agent safety & harmful actions

Insurance exposure

Agent actions

Publisher

UK AISI + Gray Swan

Publisher type

Government / standards body

License

MIT

Status

In inspect_evals (active)

Last commit

2026-10-15

Stars

704(shared repo star count — UKGovernmentBEIS/inspect_evals backs 19 entries in this catalog)

Notes

Dataset on HF: ai-safety-institute/AgentHarm.

View source on GitHub →

Wanich Risk Lab, 2026.