AI safety and guardrails

How to keep an AI application from producing harmful, biased, or policy-violating output when no attacker is involved: the ordinary failure of an ordinarily-used system. Covers what safety means for a specific application (a written list of unacceptable outcomes, not a property the model ships with), why models produce unsafe or wrong output on good-faith requests, a harm taxonomy to prioritize against, what a guardrail is and where it sits in the request pipeline, why model-level safeguards alone are never enough, the deterministic application-code controls and output moderation that do the real enforcement, designing a safe fallback (including human escalation and an emergency stop for action-taking systems), testing guardrails in both directions, why guardrails themselves cause harm when tuned carelessly, and how to detect, contain, and recover once a harmful output gets through anyway.