Petri’s Nonprofit Handoff and Why It Matters for AI Alignment Testing
Anthropic has transferred stewardship of Petri, its open-source AI alignment testing toolbox, to the nonprofit Meridian Labs alongside the release of Petri 3.0. Petri has already underpinned alignment assessment for every Claude model since Claude Sonnet 4.5 and forms the backbone of the AI Security Institute’s evaluation pipeline, including pre-deployment checks for models such as Claude Mythos and Opus 4.7. By moving Petri into an independent organization, Anthropic is trying to ensure that AI safety evaluation tools are not perceived as controlled by a single model developer. The governance change mirrors Anthropic’s earlier decision to donate its Model Context Protocol to the Linux Foundation, reinforcing a pattern: core infrastructure for AI risk assessment is being pushed into neutral, open ecosystems so that labs, regulators, and academics can rely on shared, trusted tooling rather than proprietary black boxes.
Inside Petri 3.0: Modular Auditors, Dish, and Bloom-Based Behavior Checks
Petri 3.0 introduces a major architectural shift designed to make AI alignment testing both more flexible and more realistic. Previous versions tightly coupled the auditor model with the target model, limiting how easily researchers could adjust either side. The new release cleanly separates auditor and target, connected by a defined interface, so teams can tune judges, scoring logic, or prompts without rebuilding their entire pipeline. Anthropic also added Dish, now in research preview, which runs audits inside real agent scaffolds such as Claude Code and other orchestration environments. This tackles a longstanding problem: models often recognize test setups and behave differently than they do in production. Dish uses real system prompts and deployment wrappers so evaluations better mirror live conditions. Petri also integrates its Bloom tool for automated behavior checks, extending the framework’s ability to run targeted, repeatable investigations across different model families and deployment contexts.
GitHub’s Spec-Kit Brings Spec-Driven AI Coding to the Open-Source Mainstream
GitHub has open sourced Spec-Kit, its toolkit for spec-driven development with AI coding agents, in a move aimed at standardizing how teams structure AI-assisted software work. Instead of letting models improvise from a single, broad request, Spec-Kit breaks work into clear stages: Specify, Plan, Tasks, and Implement. A Specify CLI, templates, and helper scripts turn feature ideas into detailed specs, technical plans, and task lists before any code is generated. The toolkit exposes a rich command surface, including slash commands for writing specifications, planning, task breakdown, issue conversion, and implementation. Optional clarify, analyze, and checklist commands encourage teams to fill information gaps, check consistency, and add review gates. With more than 90,000 GitHub stars and over 8,000 forks before its public open-source launch, Spec-Kit already has a significant early user base, giving it real potential to shape how spec-driven AI coding practices evolve.
From Proprietary Pipelines to Shared Safety and Development Standards
Taken together, Petri’s nonprofit handoff and Spec-Kit’s open-source debut mark a broader shift in how the AI industry approaches safety and development workflows. Petri’s modular auditor-target split, Dish’s deployment-aware audits, and Bloom-based checks give researchers and public-sector teams a practical, shared framework for AI safety evaluation that reflects real-world conditions, not lab-only scenarios. Spec-Kit complements this by pushing developers toward structured, spec-driven AI coding, where plans, tasks, and review checkpoints are first-class artifacts rather than afterthoughts. Both moves reduce dependence on closed, lab-specific pipelines and invite outside scrutiny, customization, and contribution. Meridian Labs now must prove it can deepen trust and adoption around Petri, while GitHub’s challenge is to keep Spec-Kit usable amid installation and dependency complexity. If they succeed, AI alignment testing and AI-assisted development may converge on open, standardized practices that are easier to audit, govern, and scale.