Petri is the open-source auditing framework Anthropic released in October 2025, described in this library in Petri: An open-source auditing tool to accelerate AI safety research. In it, an auditor model builds a simulated scenario from a researcher's instruction and probes a target model, and a judge model scores the transcript for misaligned behavior. This short post, published on 7 May 2026, announces two things: a third version of the tool, and the transfer of its development from Anthropic to an outside nonprofit. It reports no experimental results.
How Petri has been used
The post says Petri, developed in the Anthropic Fellows program, has been part of the alignment assessment for every Claude model since Claude Sonnet 4.5, testing for tendencies such as deception, sycophancy and cooperation with harmful requests. It also names one outside user: the UK AI Security Institute, which the post says made Petri "a major part" of how it evaluates models for their propensity to sabotage AI research (linking an April 2026 arXiv paper).
What changes in Petri 3.0
The post lists three changes:
- Adaptability. Architectural changes split the auditor and the target into separate components that can be modified independently, so the tool can be adapted to more uses.
- Realism. The post explains the motivation directly: a model can often tell from artificial details that it is being tested, and "if the model is aware it's being evaluated, the researcher is no longer able to see how the model behaves in general." An add-on called Dish makes test setups more realistic, for example by running tests with the model's real system prompt and the real scaffold (the software around the model that helps it pursue its goals) used in deployment.
- Depth. Petri is now integrated with Bloom, Anthropic's other open-source evaluation tool, which measures one chosen behavior in depth where Petri explores broadly.
Details of the release are on the Meridian Labs blog, which the post links; this note covers only what the Anthropic post states.
The transfer to Meridian Labs
Development of Petri has been handed to Meridian Labs, described as an AI evaluation nonprofit. The post compares the move to Anthropic's donation of the Model Context Protocol to the Linux Foundation and gives the reason as independence: keeping Petri separate from any AI lab "so that its results will be seen as neutral and credible by those across the industry and beyond." At Meridian Labs, Petri joins the evaluation tools Inspect and Scout in what the post calls a technology stack open to labs, independent researchers and governments.
Limitations and open questions
The post is an announcement. It gives no measurements of how much Dish reduces evaluation awareness, no comparison of Petri 3.0 with earlier versions, and no description of governance arrangements at Meridian Labs beyond the stated aim of independence. The claim that independence will make results more credible is a stated rationale, not a finding.
Research questions
- How much does Dish reduce evaluation awareness? Running the same seed instructions with and without Dish, and measuring how often target models verbalize or otherwise show awareness of being tested, would quantify the realism gain the post describes.
- Do results change when scenarios become more realistic? Comparing rates of misaligned behavior under the standard and Dish setups would show whether earlier Petri results over- or under-stated behavior in deployment.
- Does independent stewardship change who uses the tool? Tracking which organizations publish Petri-based evaluations before and after the transfer would test the post's expectation that a lab-independent tool will be more widely trusted.
- How do Petri and Bloom results compare on the same behavior? With the two integrated, measuring one behavior with both tools on the same models would show whether broad and focused evaluation agree.
How it connects
- Petri: An open-source auditing tool to accelerate AI safety research (October 2025) describes the original tool that this post updates; the post links to it.
- Introducing Bloom (December 2025) describes the tool Petri 3.0 integrates with; the post links to it.
- System Card: Claude Sonnet 4.5 (September 2025) covers the model the post names as the starting point: Petri has been part of the alignment assessment "for every Claude model since Claude Sonnet 4.5".
References
- Anthropic (2026). Donating our open-source alignment tool. 7 May 2026. https://www.anthropic.com/research/donating-open-source-petri