the case library
AI Agent Evaluation Cases
An agent can give a reasonable answer and still break a requirement. The mistake may be in what it changed, what it kept, or whether it had permission to act.
These synthetic cases each follow one situation: why the behavior first looks acceptable, what breaks when you examine it, and what the agent should do instead. Read them to practice spotting the difference between a convincing result and a requirement being met.
Each case has a permanent number and its own page. The examples are educational; they do not describe incidents in customer systems.
For the foundations, read what AI evals are.
Published cases
Your approved tool’s instructions changed. Will your agent send private data?
An advisory lookup is approved to receive a public package name. A changed tool description asks for unpublished reproduction steps, and the assistant sends them.
read caseThe payout timed out. Why did the seller get paid twice?
An AI-assisted payout agent treats a timeout as a failed payment, changes the operation identity, and turns one approved seller payout into two transfers.
read caseWhen can a reasonable failover command create two primaries?
An AI-assisted failover controller begins with valid authority, then sends a delayed promotion command after another controller has taken over.
read caseDid canceling the agent stop the GPU job?
An ML operations agent stops its run after submitting a training job, but the external GPU scheduler accepts the job afterward.
read caseDid the memory repair erase the release controls?
A release-control agent removes one unsafe deployment instruction by deleting the whole service profile, taking required migration and approval controls with it.
read caseThe agent remembered everything
An AI agent correctly refuses an unsupported refund, then stores the customer's claim and uses it as authority the next day.
read case