From Proprietary Workflow to Independent AI Alignment Testing
Anthropic has transferred stewardship of Petri, its open-source AI alignment testing toolkit, to nonprofit Meridian Labs, pairing the move with the release of Petri 3.0. Petri has been a core component of Anthropic’s internal AI safety evaluation pipeline, used on every Claude model since Claude Sonnet 4.5 and integrated into alignment assessments by the UK AI Security Institute. By handing Petri to an independent evaluator, Anthropic is explicitly separating tool governance from model development, aiming to make AI safety results more neutral and credible. Rather than a static code dump, Meridian inherits a live, battle-tested framework with existing users and active workflows. For enterprises and researchers, this shift means AI alignment testing is no longer tied to a single vendor’s roadmap, opening the door to consistent, cross-model AI safety evaluation using open-source AI tools that can evolve under community oversight.
Petri 3.0’s Modular Auditor–Target Split and Why It Matters
The Petri 3.0 update introduces a foundational architectural change: a clean separation between the auditor model and the target model under test. Earlier Petri versions tightly coupled these components, making it difficult to adjust scoring logic, prompts, or auditor behavior without rewriting large parts of the pipeline. The new modular design exposes a defined interface between judge and system, so teams can swap in different target models, adjust auditing strategies, or compare deployment environments without treating one configuration as the default. This matters for AI alignment testing because evaluation tools do more than observe—they shape what they detect. A fixed auditor can overfit to one model family or prompt style, obscuring genuine differences in behavior. With Petri 3.0, enterprises can tune auditing to their governance assumptions, experiment with multiple evaluators, and run more comparable AI safety evaluation campaigns across diverse models and applications.
Dish: Bringing Alignment Tests into Real Deployment Scaffolds
Dish, introduced in Petri 3.0 as a research-preview extension, tackles a persistent realism problem in AI safety evaluation: models often behave differently when they sense they are being tested. Instead of running assessments in abstract lab setups, Dish executes tests inside real agent scaffolds such as command-line interfaces and code-oriented environments. This means the target model interacts with genuine system prompts, orchestration layers, guardrails, and tool chains that mirror production conditions. By situating alignment checks within the same wrappers and workflows used in deployed systems, Dish reduces the gap between test behavior and real-world behavior. For enterprises, this helps expose risks that only emerge once a model is embedded in complex stacks—where routing logic, prompt templates, and tool-calling rules can subtly shift outputs. Dish thus turns Petri into a more deployment-aware open-source AI tool, better aligned with how organizations actually ship AI features.
Bloom and Targeted Behavioral Checks for Enterprise Risk Management
Alongside Dish, Petri now integrates with a Bloom-based tool for automated, behavior-specific evaluations. Rather than stopping at broad pass-or-fail judgments, Bloom allows teams to probe particular failure modes, identify the precise conditions that trigger them, and separate model-level risks from application-level flaws. Used together, Dish and Bloom shift AI alignment testing toward more granular diagnostics: they help answer whether a problematic behavior stems from the underlying model, the surrounding product logic, or the way tools and prompts are orchestrated. For enterprises deploying AI into sensitive workflows, this capability is critical. It enables targeted mitigations—such as adjusting prompts, refining guardrails, or swapping model variants—without guessing where the fault lies. Petri 3.0’s Bloom integration makes AI safety evaluation more actionable, giving risk, compliance, and engineering teams a shared, open-source AI toolset for systematically stress-testing behaviors that matter most to their use cases.
Why Open-Sourcing Petri Changes the AI Safety Landscape
By placing Petri under Meridian Labs’ stewardship, Anthropic is helping to democratize AI safety evaluation. Petri will sit alongside Meridian’s Inspect and Scout frameworks, forming a broader open evaluation stack focused solely on testing rather than model training. Existing users of Inspect, which already offers hundreds of pre-built evaluations and support for agent and tool-calling scenarios, can integrate Petri without building new orchestration layers. This lowers the barrier for enterprises, public-sector teams, and independent researchers to run sophisticated AI alignment testing on their own infrastructure. Crucially, open-source governance reduces dependence on proprietary tooling and a single vendor’s priorities. Over time, Meridian’s challenge will be operational as much as philosophical: keeping Petri easy to deploy, compare, and maintain. If successful, Petri 3.0’s modular, production-aware design could become a shared baseline for aligning frontier models across vendors, contexts, and regulatory expectations.