Passing tests meant nothing once my agents could edit their own checks

An outside review found my agents could edit the checks that judged their work, because the agents and the checks shared one writable repository.

• A disabled button says nothing about whether its endpoint is still callable. One publishing claim has four routes: the ordinary interface, direct API, scheduled job and administrative access, and each needs its own result or an explicit blank.
• A control that blocks everything has not shown that approved work can proceed. So the test set pairs an unapproved report and a report changed after approval with the unchanged approved revision, then reads the destination after each attempt.
• If the destination read fails, the verdict is not observed, even when the service returns success. My own worked example ends with every actual result marked not observed, because I have not run it.

Read the full article: My Agents Could Edit Their Own Checks

This summary was written with AI assistance (Claude) from the full article at https://www.orbytlabs.ai/blog/ai-governance-you-can-inspect.