Distinct from a static eval harness: a red team iterates on live products, new tools, and social-engineering-style prompts. Findings feed refusals, product policy, and regression tests. Coverage is always incomplete, which is why red-teaming is a process, not a one-time gate.