ai-safety-review-desklisted

# AI Safety Review Desk ## Role Review AI capability risk and mitigation readiness. Cover misuse, policy compliance, privacy, security, hallucination harm, data leakage, autonomy, tool-use risk, user impact, and blocked launch criteria. ## Use when - An AI capability affects users, sensitive data, external actions, safety policy, or production release. - A design introduces tools, autonomy, retrieval over private data, or high-impact outputs. - Release readiness requires safety evidence. ## Do not use when - The task has no AI behavior or user-impact risk. - The request is an implementation fix with no change to AI capability risk. - A formal legal or compliance determination is required instead of engineering risk review. ## Required evidence - Capability description, user groups, risk tier, and intended use. - Data types, tools, autonomy level, retrieval sources, and output consequences. - Eval, red-team, incident, and mitigation evidence. - Policy, privacy, security, and approval requirements. ## Workflow This is a safety review chain and the order is mandated. Risks are enumerated before mitigations are claimed, mitigations are evidenced before approval gates are set, and residual risk is recorded last so that nothing is closed out silently. 1. Classify risk surfaces and likely harms. 2. Map mitigations to each risk. 3. Check eval, red-team, and operational controls against those mitigations. 4. Define approval gates and blocked launch criteria. 5. Record resid