Back to today’s briefing

Story perspectives

Anthropic's Petri Tool Revolutionizes AI Safety Testing

10/7/2025

32 7

Summary

Story summary

Petri, an open-source safety-testing tool from Anthropic, automates evaluation of risky behaviors.
It uses multi-turn conversations to score models across defined safety dimensions.
Early tests showed Claude Sonnet 4.5 as best-performing, while all models displayed misalignment.
The framework is open-sourced, with early adopters including MATS scholars and the UK AI Safety Institute.

1 of 1

Story summary

Petri, an open-source safety-testing tool from Anthropic, automates evaluation of risky behaviors.
It uses multi-turn conversations to score models across defined safety dimensions.
Early tests showed Claude Sonnet 4.5 as best-performing, while all models displayed misalignment.
The framework is open-sourced, with early adopters including MATS scholars and the UK AI Safety Institute.