Guardrails

Checks around the model that block, rewrite or escalate inputs and outputs that break policy.

Evolving: the basics hold, details shift. Reviewed 17 Oct 2026.

Key ideas

• Input checks: topic scope, PII detection, injection and jailbreak classifiers.
• Output checks: schema validation, grounding checks, toxicity, PII leakage, policy rules.
• Options: Bedrock Guardrails, Azure AI Content Safety, NeMo Guardrails, Llama Guard, custom classifiers.
• Run cheap checks in parallel with generation to avoid adding latency.

Common mistakes

• Over-blocking legitimate requests; measure false positives as carefully as misses.