Agent Red-Team Evaluations

Evaluate agentic AI under adversarial pressure. Learn agent-eval suites and how to harden them with adversarial inputs (METR-style autonomy, SWE-bench under injection, AgentBench under hostile environment), red-team-specific harnesses, scoring rubrics for partial compromise, the link to capability evaluations, and operational guard rails (sandboxes, max-steps, max-cost, human checkpoints).