Guardrails
Guardrails are multi-layered protection by application layer:
- input — relevance classifier, jailbreak and injection detection, moderation, rules;
- execution — tool risk assessment before invocation;
- output — PII filtering, brand compliance control.
One guardrail is not enough — a combination is needed. Plus human intervention (human in the loop) at critical points.
Related: Prompt injection, Human in the loop, Deadly triad