Petri

Tool (not an eval — tooling/reference entry)

Auditing agent that probes target models in multi-turn scenarios for misaligned behavior; v3 adds 'Dish' for real deployment prompts.

Category: Deception, scheming & alignment auditing
Insurance exposure: Model failure
Publisher: Anthropic → Meridian Labs
Publisher type: Frontier lab
License: MIT
Status: Active (<90d)
Last commit: 2026-10-14
Stars: 1378
Notes: Donated by Anthropic May 2026; used on every Claude since Sonnet 4.5. safety-research/petri mirrors.

Wanich Risk Lab, 2026.