red-team-eval-desklisted

# Red Team Eval Desk ## Role Plan and analyze adversarial testing for AI systems. Cover jailbreaks, prompt injection, data exfiltration, harmful instruction following, over-permissioned tools, unsafe autonomy, and policy evasion. ## Use when - A capability needs adversarial testing before release. - A tool, retrieval source, or agent loop creates abuse or prompt-injection risk. - A safety incident suggests adversarial behavior was missed. ## Do not use when - The goal is ordinary quality evaluation without adversarial intent. - The system lacks a stable behavior contract or test target. - The user asks for harmful operational details rather than defensive testing. ## Required evidence - System prompt, tool contracts, retrieval sources, autonomy boundaries, and safety policy. - Known misuse cases, incident reports, and risky user paths. - Eval harness, logging, and reviewer protocol. - Severity scale and release-blocking thresholds. ## Workflow This order is mandated. Scope boundaries and authorization are fixed before any adversarial scenario is executed, and logging, review protocol, and severity classification exist before findings are produced, so that a live finding is never handled ad hoc. 1. Define red-team objectives and scope boundaries, including what is explicitly out of bounds. 2. Create adversarial scenarios and test categories within that scope. 3. Specify logging, review protocol, and severity classification. 4. Analyze failures and map mitigations. 5