inspect_evals

Framework / harness (not an eval — tooling/reference entry)

Library of 100+ ready-to-run evals incl. Cybench, CyberSecEval 2-4, CVE-Bench, CyberGym, AgentHarm, AgentDojo, WMDP, GDM dangerous-capability tasks.

Category: Eval frameworks & harnesses
Insurance exposure: All
Publisher: UK AISI + community
Publisher type: Government / standards body
License: MIT
Status: Active (<90d)
Last commit: 2026-10-15
Stars: 704(shared repo star count — UKGovernmentBEIS/inspect_evals backs 19 entries in this catalog)
Notes: Single best entry point. Individual risk evals in it are listed separately below.

Wanich Risk Lab, 2026.