Anthropic hands its open-source alignment testing tool Petri to Meridian Labs

In short

Anthropic says it has released a third version of Petri, an open-source tool that tests AI models for tendencies such as deception and sycophancy, which means telling people what they want to hear. Anthropic also says it has handed development of Petri to Meridian Labs, an AI evaluation nonprofit.

What happened

Anthropic reported the changes on May 7, 2026. Petri launched in October 2025 as a set of alignment tests that can be used on any large language model. Alignment means whether an AI system behaves in the ways its developers intend. Anthropic says Petri was developed through its Anthropic Fellows program.

Anthropic says Petri has been part of its alignment assessment for every Claude model since Claude Sonnet 4.5. In these tests, a separate "auditor" model plays out simulated scenarios with the model being tested. A "judge" model then scores the transcripts for misaligned behaviour.

Anthropic says outside groups also use it. It says the UK's AI Security Institute (AISI) made Petri a major part of how it evaluates models for their tendency to sabotage AI research.

Anthropic describes three main changes in Petri 3.0:

• Adaptability: The auditor model and the target model are now separate components, so users can adjust each one on its own.
• Realism: Anthropic says a model can often tell from artificial details that it is being tested, and then the test may not show how it normally behaves. An add-on called "Dish" is meant to make tests more realistic. It uses the model's real system prompt and the real "scaffold," which is the software that wraps around a model to help it meet its goals.
• Depth: Petri is now linked with Bloom, another Anthropic open-source tool. Anthropic says Bloom can assess specific chosen behaviours in more depth, while Petri covers a wider range.

What was done about it

This report describes a tool update, not a specific incident. Anthropic says it has handed Petri's development to Meridian Labs. It compares the move to its donation of the Model Context Protocol to the Linux Foundation. Anthropic says the aim is to keep Petri independent of any AI lab, so that its results are seen as neutral and credible. Petri now sits alongside Meridian Labs tools called Inspect and Scout. More detail on Petri 3.0 is on the Meridian Labs blog, according to Anthropic.

Why it matters

Anthropic says that reliable tests of AI model behaviour matter more than ever. It argues that an independent home for Petri will help labs, independent researchers and governments trust the results. The report does not give test results for Petri 3.0 or say how well the realism changes work. Readers can find more in the original report.