Anthropic Opens Petri AI Safety Auditing Framework and 111 Seed Instructions

Anthropic has released Petri, an open-source framework designed to automate parts of AI alignment auditing. The Parallel Exploration Tool for Risky Interactions runs parallel agent conversations against target models, scores the resulting transcripts, and gives researchers a way to inspect potentially misaligned behavior. Alongside the framework, Anthropic published an initial set of 111 seed instructions intended to help users begin audits.

The release matters because it makes a concrete safety-evaluation workflow available beyond Anthropic's own research environment. As described in Anthropic's official Petri release, the tool is built to test behaviors including deception, sycophancy, whistleblowing, reward hacking, self-preservation, and other forms of misalignment. The prompts are explicitly presented as evolving starting points rather than finished benchmarks.

What Anthropic's Petri release includes

Petri is an auditing framework, not a claim that any individual test can settle whether a model is safe. Its value is in creating a repeatable process for exploring risky interactions, recording what happens, and reviewing the evidence. Anthropic says the framework can run across major model APIs and includes an automated judge and a transcript viewer for analysis.

Framework, prompts, and analysis workflow

The public release combines several pieces that are useful together:

• The Petri framework, available through Anthropic's GitHub repository, for conducting automated alignment audits.
• 111 seed instructions that guide initial tests of diverse model behaviors associated with misalignment.
• Parallel agent conversations that generate transcripts for a target model under audit.
• An automated judge and transcript viewer to support scoring and human review of those interactions.

This structure can reduce the practical barrier to running exploratory evaluations. Rather than starting with a blank test plan, teams can adapt an existing seed instruction, run it against a model accessible through an API, and inspect the resulting conversation. That does not remove the need for expert judgment. It does, however, make the methodology and starting materials more visible and reusable.

The final row is related to Anthropic's broader research practices, but it is a separate effort from Petri. The Anthropic Economic Index report states that its underlying data was open-sourced so researchers and policymakers could further study AI's economic effects. Its dataset covers task-level usage patterns, transcripts, and methodology associated with that report. It should not be treated as part of the Petri code and seed-instruction release.

Why the seeds are useful, but not definitive

Open-sourcing seed instructions is significant for reproducibility because other researchers can see the initial framing of an audit and modify it for their own tests. It also enables more meaningful comparisons when different teams use related starting materials against different models or APIs.

At the same time, Anthropic describes the seeds as rough and evolving. That limitation is important. A seed instruction is not a comprehensive benchmark, a universal safety standard, or a guarantee that an observed behavior will generalize. Results can vary with the target model, model version, system configuration, API access, and the details of the audit setup. Researchers using Petri should document those conditions carefully if they want others to interpret or reproduce their findings.

Reproducibility and governance implications

The Petri release offers a more tangible form of openness than a technical description alone. Publishing code, initial prompts, and a workflow gives third parties materials they can inspect, adapt, and challenge. For AI safety work, that can strengthen external scrutiny by moving some evaluation practice from high-level discussion into testable implementation.

The separate open-data commitment in Anthropic's Economic Index research supports a similar principle in a different domain. Making underlying data available can allow independent analysis of reported patterns and methods, particularly where the findings may inform policy discussions. Together, the releases point to an approach that combines safety engineering resources with research data intended for wider examination.

There are still practical boundaries. Open tooling does not ensure identical results when model access differs across providers or when models change over time. The availability of seed instructions also does not answer every question about how sensitive evaluation content should be shared or deployed. Teams should review the applicable repository documentation and licensing terms before incorporating Petri or its materials into internal evaluation pipelines.

For developers and safety teams, the immediate opportunity is to treat Petri as a structured starting point for internal red-teaming and comparative testing. A useful deployment path is to select relevant seed instructions, adapt them to a clearly defined risk scenario, retain transcripts and configuration details, and combine automated scoring with expert review. The framework can support investigation, but governance decisions should remain grounded in an organization's own risk thresholds and accountability processes.

As AI systems become part of customer-facing and operational workflows, businesses need evidence about how those systems are represented and evaluated across AI environments. Scalevise helps teams assess their presence in AI-driven answers, identify visibility gaps, and turn findings into a practical improvement plan through its AI Visibility and GEO Checker. A disciplined visibility baseline can help leaders connect AI adoption decisions with measurable market discovery and governance priorities. Start an AI Visibility scan.

Frequently Asked Questions

What is Anthropic Petri?

Petri is Anthropic's open-source Parallel Exploration Tool for Risky Interactions, a framework that automates parts of alignment auditing through parallel agent conversations, transcript scoring, and analysis tools.

How many seed instructions did Anthropic release with Petri?

Anthropic released an initial set of 111 seed instructions. They are intended as adaptable starting points for testing a range of potentially misaligned model behaviors.

Which behaviors can Petri's seed instructions help examine?

Anthropic says the seeds can be used to test behaviors including deception, sycophancy, whistleblowing, reward hacking, self-preservation, and other misaligned behaviors.

Is the open data mentioned by Anthropic part of the Petri release?

No. The open data relates to separate Anthropic research, including the Anthropic Economic Index report, whose underlying data was released to support independent analysis.

Conclusion

Anthropic's Petri release provides an open-source framework and an initial prompt set for exploring alignment risks in a more repeatable way. By separating reusable audit tooling from finished safety judgments, the project gives researchers and practitioners materials they can inspect, extend, and test across available models. Its value will depend on careful adaptation, transparent documentation, and rigorous human interpretation of the results.