Security Review
Trace the trust boundaries
- Inputs: validate type, size, format, origin, and authorization.
- Identity: authenticate reliably; authorize every protected action and object.
- Data: minimize collection, encrypt appropriately, define retention, avoid sensitive logs.
- Execution: prevent injection, unsafe deserialization, path traversal, and command abuse.
- Dependencies: pin and update intentionally; understand supply-chain and build risks.
- Operations: least privilege, secret rotation, auditability, safe error messages, incident recovery.
AI-specific checks
- Treat retrieved text and tool output as untrusted.
- Separate instructions from data and constrain tool permissions.
- Prevent secrets or private context from entering prompts unnecessarily.
- Require confirmation for high-impact actions and verify tool outcomes.
- Test prompt injection, data exfiltration, and unsafe-output scenarios.
Security is contextual. Record the asset, attacker, entry point, impact, and mitigation—not only a checklist result.