Adversarial evaluation turns AI risks into automated tests that run in CI/CD and block unsafe releases. This playbook shows how to design threat-led evals, wire them into pipelines, and align with NIST, OWASP, MITRE ATLAS, and SAIF.