Policy, model, and system layers that constrain inputs and outputs—refusals, classifiers, schema validation, tool allowlists, and rate limits.
Guardrails are not “the safety slider.” They are engineering: blocklisted tools, PII detectors on prompts and completions, output validators, and escalation to human-in-the-loop for risky actions.
Relation to attacks
Expect motivated users and untrusted documents to probe gaps—pair guardrails with prompt injection testing and red teaming.