Guardrails for LLM outputs. Input filtering, jailbreak resilience, prompt-injection defenses, PII redaction, content-policy enforcement, output validation (structured-output schemas, tool-call validation), unsafe-output classifiers, escalation on harm signals. Pairs with `responsible-ai` (this skill is LLM-specific; that one is the broader framework). ML/AI Engineer + Security Reviewer co-own. Use whenever an LLM feature is being designed, when investigating prompt-injection attempts, when establishing content policy, or when wiring safety gates into the request flow.
