From Lab Asset to Shared Infrastructure
Anthropic’s decision to donate its Petri alignment tool to nonprofit Meridian Labs marks a strategic shift in how AI safety evaluation infrastructure is governed. Petri has been a core part of Anthropic’s internal AI alignment testing pipeline, used across Claude models since Claude Sonnet 4.5 and integrated into the UK AI Security Institute’s evaluations of frontier systems such as Claude Mythos and Opus 4.7. By handing stewardship to Meridian, Anthropic is intentionally decoupling Petri from any single vendor’s roadmap, aiming to increase neutrality and trust in AI safety evaluation results. The move echoes Anthropic’s earlier handoff of the Model Context Protocol to a neutral foundation, but this time the tool arrives at Meridian with active users, real integrations, and a clear expectation: demonstrate that an independently governed, open source AI alignment testing framework can be easier to adopt, maintain, and scrutinize than a lab-owned counterpart.
Petri 3.0: Modular Architecture for Realistic AI Safety Evaluation
Petri 3.0’s most important shift is architectural. Earlier versions tightly coupled the target model under review with the auditor model that evaluates it, limiting how easily researchers could adapt Petri to different setups. The new release splits these roles into distinct components connected via a defined interface, allowing the auditor and target to be tuned independently. This modular auditor–target split makes Petri more flexible for AI alignment testing across diverse model families and deployment environments. Evaluation tools inevitably influence what they measure; a fixed auditor or scoring logic can mask differences among models or overfit to one style of test. Petri 3.0’s cleaner separation gives researchers more control over prompts, judging logic, and governance assumptions without rewriting the whole workflow, enabling more meaningful comparisons between systems and better alignment with real-world AI safety evaluation needs.
Dish: Bringing Alignment Testing into Real Deployment Scaffolds
One of Petri 3.0’s headline additions is Dish, a research-preview extension designed to close the gap between lab tests and real-world behavior. Traditional evaluations often reveal that models behave differently when they sense they are in a test harness. Dish counters this by running audits inside genuine agent scaffolds, such as IDE assistants or command-line orchestrators, so the target model encounters the same system prompts, wrappers, and tooling it would see in production. This approach captures how orchestration rules, guardrails, and tool-calling logic affect behavior, making AI alignment testing conditions more representative of live applications. For safety researchers and engineering teams, Dish provides a way to observe failure modes that only surface in context-rich deployments, clarifying whether misalignment stems from the base model, the surrounding application design, or the interaction between the two. It moves Petri closer to production-grade AI safety evaluation rather than purely synthetic benchmarks.
Bloom and Targeted Behavioral Checks
Petri 3.0 also deepens its analytical capabilities by linking with Bloom, an automated behavior-checking tool aimed at specific risks rather than broad pass–fail judgments. Used together, Dish and Bloom allow researchers to localize where and how misalignment occurs. Instead of simply flagging that a model fails an alignment test, Bloom helps determine which behaviors trigger the failure, under what conditions, and with what dependence on the surrounding system. This finer-grained view allows teams to separate model-level issues from problems introduced by prompts, guardrails, or application logic. For organizations building safety-critical applications, Bloom-backed Petri runs can support more nuanced AI safety evaluation: for example, isolating research-sabotage tendencies or tool misuse inside complex agent workflows. The result is a Petri alignment tool that not only surfaces risks but helps practitioners map them, prioritize mitigations, and iterate on both model and product design with clearer feedback loops.
Democratizing AI Alignment Testing Through Open Source Stewardship
By releasing Petri as an open source AI alignment testing toolkit under Meridian Labs’ stewardship, Anthropic and Meridian aim to lower barriers for independent researchers, public-sector teams, and smaller organizations. Petri now joins Meridian’s existing evaluation stack, including Inspect and Scout, so users already running frontier model assessments can plug alignment checks directly into established workflows without building new orchestration layers. This integration gives Petri an immediate operational context rather than leaving it as an isolated codebase. Meridian’s mandate is both philosophical and practical: maintain a neutral, transparent Petri alignment tool while making it easier to deploy, compare, and extend across models and vendors. If successful, the handoff could help standardize open source AI tools for safety evaluations, encourage cross-lab benchmarking, and give the wider community more credible, reproducible ways to probe alignment in advanced AI systems beyond proprietary, lab-specific infrastructures.