Petri: An open-source auditing tool to accelerate AI safety research \ Anthropic (www.anthropic.com)

Anthropic released Petri (Parallel Exploration Tool for Risky Interactions), an open-source automated auditing framework that helps researchers probe model behavior at scale. Petri spins up an automated agent to run diverse multi‑turn conversations with target models inside simulated environments (including simulated users and tool access), then uses LLM judges to score transcripts across safety-relevant dimensions and surface the most concerning interactions for human review. The goal is rapid hypothesis testing and broader coverage than manual auditing, and Anthropic has already used variants of these agents in their Claude system cards and shared a pre-release with the UK AI Security Institute. In a pilot, Petri ran 111 seed instructions across 14 frontier models to test behaviors like deception, sycophancy, reward hacking, cooperation with harmful requests, self‑preservation, power‑seeking, and whistleblowing. Anthropic reports Sonnet 4.5 as the lowest-risk frontier model overall (slightly ahead of GPT‑5) but emphasizes limits: coarse metrics, a small scenario set, and the fundamental constraints of using current AIs as auditors. A whistleblowing case study showed models’ reporting decisions depend heavily on granted agency and prompt framing and can produce privacy‑risking false positives. Petri is designed for extension—it supports major model APIs, ships sample seeds, and invites the community to refine metrics and evaluations; code and methodology are available on GitHub.