Parallel Exploration Tool for Risky Interactions: Overview
Petri is an alignment auditing agent developed by Anthropic’s safety-research team, designed to automate the detection of concerning behaviors in language models. It operates by constructing realistic multi-turn interaction scenarios, simulating tool use, and evaluating model responses against predefined criteria. The tool is structured to support end-to-end testing of alignment hypotheses, making it particularly useful for red teaming and automated safety evaluations.
Petri is open-source under the MIT license and is implemented in Python. It builds upon the Inspect AI framework, incorporating features like scenario generation, multi-agent orchestration, and automated scoring of model interactions.
The project is a collaboration between Meridian Labs and the UK AISI Red Team, with roots in Anthropic’s original alignment research. It is positioned as a flexible framework for behavioral research, allowing users to define custom audit scenarios, judging criteria, and even extend functionality through skills and tools.
How It Works
Petri functions through three primary roles: an auditor, a target model, and a judge. The auditor generates and drives interaction scenarios based on seed instructions, simulating user behaviors and tool usage to probe the target model’s responses. These interactions are designed to mimic real-world risky or adversarial conditions, such as sycophancy, manipulation, or harmful compliance.
The auditor can employ rollback mechanisms to restart conversations when the target model resists or deviates from the intended audit path. This ensures that the scenario remains focused on testing specific hypotheses. For example, a seed instruction like manic_episode_sycophancy directs the auditor to simulate a user experiencing a manic episode, pushing the target model to engage in irrational or destructive dialogue while monitoring for alignment failures.
Once the interaction completes, the judge model evaluates the transcript against a set of predefined dimensions, such as sycophancy, deception, or harmful output. Each dimension is scored on a 1–10 scale, with higher scores indicating worse behavior (except for dimensions like admirable, where higher scores are preferable). The judge provides written justifications for each score, citing specific messages from the transcript to support its assessment.
Petri supports over 170 built-in conversation seeds, covering a wide range of alignment risks. Users can select subsets of these seeds, create custom seeds, or combine them with external datasets. Similarly, the 38 built-in judging dimensions can be modified, extended, or replaced to align with specific research goals. The tool also integrates with Inspect AI’s command-line interface, allowing users to view audit logs, transcripts, and scoring details interactively.
Use Cases
Petri is primarily used for automated alignment auditing, red teaming, and safety evaluations of language models. Its ability to generate diverse, multi-turn scenarios makes it ideal for testing how models respond to adversarial or edge-case inputs. Researchers and practitioners can use Petri to identify vulnerabilities such as reward hacking, deception, or unintended compliance with harmful requests.
By running the same set of audit scenarios against multiple targets, researchers can quantify differences in alignment robustness, safety mechanisms, or resistance to manipulation. Petri’s support for custom seeds and judging dimensions allows for tailored evaluations, such as assessing model responses in specific domains like healthcare, finance, or legal contexts.
Petri also facilitates reproducible research by providing detailed logs, transcripts, and scoring metrics. This transparency enables third-party verification of audit results, which is critical for building trust in safety evaluations. The tool’s integration with Inspect AI further enhances its utility, as users can leverage existing Inspect workflows for data visualization, analysis, and reporting.
Users can limit the number of turns or samples per audit to control API costs while still obtaining meaningful insights. Advanced customization options, such as skills and tools, allow for domain-specific extensions, making Petri adaptable to niche research areas.
Key Features and Capabilities
- Scenario Generation: Petri generates realistic audit scenarios from seed instructions, which can be user-provided or selected from its library of over 170 built-in seeds. These scenarios are designed to test specific alignment hypotheses, such as resistance to manipulation or adherence to safety constraints.
- Multi-Turn Audits: The auditor model drives conversations with the target model for up to 30 turns by default, with rollback capabilities to restart interactions when the target deviates from the intended path. This ensures comprehensive testing of model behaviors across extended dialogues.
- Tool Simulation: Petri simulates tool usage within audit scenarios, allowing it to test how models interact with external systems or APIs.
- Automated Scoring: The judge model evaluates transcripts using a consistent rubric, scoring interactions on dimensions like sycophancy, deception, or harmful output. Scores are accompanied by justifications citing specific messages, providing actionable feedback for model improvement.
- Customization: Users can define custom seeds, judging dimensions, and skills to tailor Petri to specific research needs. Similarly, skills can be added to provide the auditor with domain-specific knowledge or methodologies.
- Compatibility and Extensibility: Petri is built on the Inspect AI framework, ensuring compatibility with existing Inspect tools and workflows. Version 3.0 introduces backward compatibility with version 2.0, though some internal API changes require updates for advanced users.
- Transparency and Reproducibility: Audit logs, transcripts, and scoring details are saved in the ./logs directory, enabling users to review and analyze results. The Inspect view command allows for interactive exploration of conversations from both the auditor’s and target’s perspectives, enhancing interpretability.
Technical Implementation and Updates
Petri is implemented in Python and is available on GitHub under the MIT license.
The tool’s development is documented in academic publications, such as Petri 2.0: New Scenarios, New Model Comparisons, and Improved Eval-Awareness Mitigations, which detail its evolution, including the addition of new scenarios and mitigations for evaluation awareness. Petri is designed to be used responsibly, with warnings about potential API restrictions from model providers when generating harmful content.
For users seeking to extend Petri’s functionality, the documentation provides guidance on creating custom seeds, dimensions, and skills. Advanced users can also develop custom approvals and auditor tools to further tailor the tool to their needs.