AI Red Teaming Methodology

Red-teaming an AI system is not a classic pentest — the failures are behavioural and probabilistic, so you hunt for harmful outputs, jailbreaks, injection and leakage instead of CVEs. This is the full method: scope the harms, map the surface, generate and mutate attacks, run the automated attacker-target-judge loop at scale, and wire the whole thing into CI with PyRIT, Garak, and Promptfoo so safety does not silently regress on the next model bump.