Architecture & Patterns: Security + Guardrails¶

Production Patterns¶

Architecture View¶

graph LR
  A[User input] --> B[Input policy check]
  B --> C[Model response]
  C --> D[Output redaction]
  D --> E[Tool allowlist]
  E --> F[Safe execution]

Common Failure Modes¶

Relying on prompt instructions without runtime checks.
Logging raw sensitive outputs before redaction.
Broad tool permissions with no path or command scope.

Interview Q&A¶

Q: Interview Q: Why is defense-in-depth necessary for agent systems?

Any single guardrail can fail under adversarial input. Layered controls reduce risk and provide fallback containment.

Q: Interview Q: How do you evaluate guardrail quality over time?

Track blocked and allowed outcomes against labeled test sets, then monitor false positives and false negatives per policy version.