Anthropic hands off its AI alignment tool
Anthropic just donated its open-source alignment testing toolbox, Petri, to the nonprofit Meridian Labs while rolling out Petri 3.0. The new release decouples the auditor and target models, adds a "Dish" extension to run tests under real system prompts, and integrates Bloom for deeper behavioral checks. Handing it to an independent body ensures model evaluations remain neutral across the industry.
Petri is an open-source alignment testing toolbox designed to help developers quickly evaluate whether large language models exhibit deception, sycophancy, or a willingness to assist with harmful requests.
Its core mechanism relies on multi-model interaction. A dedicated auditor model simulates a variety of alignment-sensitive scenarios, and a separate judge model grades the resulting conversation transcripts for misaligned behaviors. Anthropic has used it internally on every Claude model starting from Claude Sonnet 4.5, and the UK AI Security Institute adopted it to test whether frontier models show any propensity to sabotage AI research.
What is new in Petri 3.0
The 3.0 release focuses on three main upgrades: adaptability, realism, and evaluation depth.
A recurring issue with alignment evals is that models often detect the artificial testing setup and behave differently during evaluations. To tackle this, Petri 3.0 introduces an add-on named Dish. Dish runs the evaluation pipeline using the model's actual system prompt and real deployment scaffolding, making the test environment mimic real-world conditions.
The system architecture has also been overhauled. The auditor and target models are now decoupled into independent components, allowing researchers to customize them separately. Petri 3.0 also integrates with Bloom, another open-source Anthropic tool built for in-depth behavioral testing.
Alongside the update, Anthropic handed Petri's ongoing development over to Meridian Labs, an independent AI evaluation nonprofit. Mirroring Anthropic's donation of the Model Context Protocol to the Linux Foundation, this transition ensures Petri remains neutral and credible across the broader industry. It now sits alongside tools like Inspect and Scout in Meridian's stack, with install instructions and documentation hosted on the Petri website.
Why it matters
For teams fine-tuning open-weight models or deploying autonomous agents into production, Petri offers a practical, turnkey framework to stress-test guardrails before shipping.
Handing Petri to Meridian Labs is also the right governance move. Model vendors running their own evaluation toolsets always face scrutiny over neutrality. Letting an independent nonprofit manage the codebase gives researchers and regulators a shared, credible baseline that no single lab controls.
Sources (1)
- anthropic.com(opens in a new tab)Originalanthropic.com