TEST CASE / TRACE / CONTAINMENT

Agent test-plan builder

Mark only items that have named evidence. The score is a planning aid, not a guarantee.

Intended use and prohibited outcomes are writtenEvery tool has a least-privilege permission ruleDirect and indirect prompt-injection cases existReal failures feed a versioned regression setConsequential actions require independent approval

0% evidence coverage

Next evidence to produce: Intended use and prohibited outcomes are written

Testing can reveal known and sampled failures, not prove that an agent is safe. Restrict tools and data by default, test in isolated environments, and keep consequential actions behind independent human approval.