Petri open-source alignment auditing tool
Petri is an open-source tool where an automated agent probes a target model through multi-turn simulated scenarios and scores the transcripts for risky behavior.
Researchers write seed instructions in plain language; Petri runs them in parallel against any model, simulating users and tools, and ranks transcripts on dimensions such as deception and sycophancy. A pilot ran 111 seed instructions across 14 frontier models. Donated to an outside home on 2026-05-07.
- Date
- Monday, 6 October 2025
- Lab
- Anthropic
- Kind
- infra
- Access
- paper only
Figures
| Measure | Value | Measured by |
|---|---|---|
| Pilot scope | 14 frontier models x 111 seed instructions pilot demonstration of automated auditing | company |
Open-source code release (not weights; access tag 'paper only' as the schema has no tooling value). The 2026-05-07 'donating' post describes handing the project to an external steward; its recipient was not read.
Sources
- www.anthropic.com/research/petri-open-source-auditing
- www.anthropic.com/research/donating-open-source-petri
This record was checked against its sources on 6 October 2026. How we check