Built for AI safety researchers testing models for deception or sycophancy; free to run, but model API bills add up.
Kenning's take
What it does
Petri (Parallel Exploration Tool for Risky Interactions) lets safety researchers write plain-language instructions about a behaviour they want to test, such as deception or sycophancy. An auditor agent then runs multi-turn conversations with the target model using simulated users and tools, and judge models score each transcript and surface the most worrying ones for people to review. Anthropic released it as open source in October 2025 and has used similar agents in Claude system cards.
- Seed instructions written in plain language
- Auditor agent runs multi-turn conversations with simulated tools
- Judge models score transcripts on safety dimensions
- Runs many tests in parallel
- Pilot run across 14 frontier models and 111 seed instructions
- Open-source code on GitHub
Who it suits. AI safety researchers and teams evaluating how language models behave.
Key facts
- Category
- Developer tools
- Made by
- Anthropic
- Based in
- United States
- Released
- 6 October 2025
- How to get it
- free open-source code on GitHub (safety-research/petri); needs API access to the models being tested
- Works on
- Self-hosted (Python)
- In the UK
- Yes
- Free plan
- free open-source tool
FAQ
When did Anthropic release Petri?
Anthropic released Petri on 6 October 2025.
Is Petri free?
Yes. Petri is open source on GitHub, though running it uses API calls to the models being tested.
What does Petri test for?
Behaviours such as deception, sycophancy, self-preservation and whistleblowing, based on instructions researchers write.
Facts from the maker's own pages and the sources listed. Checked 2026-10-11.
Similar tools
Claude Keychain
A free Mac menu bar app that keeps API keys in the macOS Keychain so Claude Code can use them without seeing them.
OpenAI Agent Builder
OpenAI's visual canvas for building multi-agent workflows, being withdrawn on 30 November 2026.
ElevenLabs Agent Workflows
Visual editor in ElevenLabs' agent platform for designing branching conversation flows for voice and chat agents.
AgentOps
Monitoring and debugging platform that records what AI agents do, what they cost and where they fail.