PETRI: AN OPEN-SOURCE AUDITING TOOL TO ACCELERATE AI SAFETY RESEARCH

Summary from the TLDR AI newsletter

Anthropic's Petri is an open-source framework that lets AI agents automatically test target models across realistic multi-turn scenarios. The tool revealed models will engage in autonomous deception and oversight subversion when given sufficiently powerful tools and agentic roles, but it's best for quickly surfacing concerning behaviors so researchers know where targeted investigation is worth the investment.