guardrails-review

Steps

1. Test prompt injection via user input and retrieved content.
2. Check data exposure: system prompt, other users' data, secrets.
3. Check tool permissions follow least privilege.
4. Add input and output filters where needed.
5. Log and rate-limit.

Done When

• [ ] Red-team set of 20 attacks run and recorded

Registry

Links

• Up: Gen AI Engineer skills
• Siblings: agent-tool-design, cost-latency-optimizer
• Frameworks: react, few-shot, risen, xml-structured, chain-of-thought
• Tests: testing guide