← Back

Common Guardrails

·Bryan Lai
TechnicalGuardrailsLLM

Common Guardrails

Guardrails fail when teams treat them as one filter.

Good systems layer checks before, during, and after the model responds.

  1. Relevance: Is the answer still on topic?
  2. Safety: Is the user trying to jailbreak the system or inject instructions?
  3. PII: Did the model expose personally identifiable information?
  4. Moderation: Is the input or output hateful, violent, sexual, or otherwise unsafe?
  5. Tool risk: Is the tool call safe to run now, or should a human approve it first?
  6. Rules: Are there known bad patterns to block with regex, blocklists, or SQL checks?
  7. Output: Would this response harm the user, the business, or the brand?

Heuristics

  1. Focus on data privacy and content safety.
  2. Add guardrails after real failures, not imaginary ones.
  3. Keep the user experience usable.
  4. Tighten the rules as the agent gets more power.