Model the whole attack path
Identify valuable assets, entry points, trust boundaries, privileges, and possible effects. Combine adversarial content with tool access, caches, and integrations in testing. A filter that blocks a few phrases is not proof that the path is closed.
Make risk decisions reviewable
Assign an owner, define unacceptable outcomes, document assumptions, and select controls proportionate to impact. Keep evidence of evaluation and exceptions. Model cards and vendor assurances can inform review, but do not replace assessment of your particular deployment.
Prepare to contain and recover
Define how to disable actions, revoke access, preserve appropriate evidence, notify responsible people, and restore a known configuration. Test the runbook. Reassess when models, data, tools, or permissions change; a past approval is not permanent proof of safety.
A worked example
Run regression and adversarial evaluations in a controlled environment, inspect permission boundaries, and stage the update with rollback criteria. Keep high-impact actions constrained until the evidence supports release.
The point: The deployed system includes the model version and its surrounding controls.
Before you take the exam
Explain these concepts in your own words, work through a different example, and describe why the tempting alternatives fail. Use the references to go deeper.