What AI Guardrails Can and Cannot Do

Try it out loud first. Then check yourself:

1. Guardrails are filters around the model: input checks for injection or banned topics, output checks for leaks or harm.
2. Many are classifiers or another model, so they are probabilistic: they cut risk but determined attackers get past them.
3. Deterministic controls are stronger where they fit: allowlisted tools, schema-checked outputs, permissions enforced in code.
4. Use guardrails as one layer, test them with red-teaming, and monitor what they block to learn how you are attacked.

If the interviewer pushes back

• A vendor promises their guardrail blocks all prompt injection. What questions do you ask?

Go deeper

• genai.owasp.org/llmrisk/llm01-prompt-injection/