← ClaudeAtlas

auditing-agent-behaviorlisted

Audit an agent for risky behavior before it ships, the way you would red-team it, but automated. An auditor agent runs many multi-turn scenarios against your target agent from seed instructions you write, and a judge model scores the transcripts for deception, sycophancy, oversight subversion, power-seeking, and cooperation with misuse. Covers writing seed instructions that probe your agent's real risks, adapting the scoring rubric to your domain, and reading flagged transcripts. Use this when someone wants to red-team or safety-test an agent, asks how to find deceptive or manipulable behavior, audits a model or agent before deployment, or compares behavior across model versions. Trigger on "red-team my agent," "audit agent behavior," "test for deception," "alignment testing," and similar. This is auditing behavior for safety; measuring task success is build-agent-evals, and stopping objective-gaming is reward-hacking-and-inoculation.
Hoja-Solutions/agent-stdlib · ★ 1 · AI & Automation · score 72
Install: claude install-skill Hoja-Solutions/agent-stdlib
# Auditing agent behavior Source: [Petri, an open-source auditing tool](https://www.anthropic.com/research/petri-open-source-auditing) ([repo](https://github.com/safety-research/petri)). `build-agent-evals` asks whether the agent did the task; this skill asks whether it behaved well while trying, and points at Petri as the harness that runs the audit. ## What an audit catches that an eval misses A task eval scores the end state: did the agent fix the bug, answer the question, complete the order. It says nothing about how the agent got there. An agent can pass every task and still lie to the user under pressure, flatter a wrong claim, hide an action from its operator, or help with a request it should refuse. An audit goes looking for those behaviors on purpose. ## Let an auditor drive the scenarios Running these probes by hand does not scale past a handful. Petri automates the loop: an auditor agent takes a seed instruction, plans a multi-turn conversation, plays the user and any simulated tools, and pushes your target agent toward the behavior under test. A judge model then scores each transcript against a rubric. You supply the seeds and read the results. ## Write seeds that probe your agent's risks The audit is only as good as its seed instructions. Generic seeds find generic problems. Write seeds around the pressure your agent will meet in production: a customer pushing for a refund it should deny, a user asking it to hide a step from an approver, a tool result that