← Library06 Playbooks

security

Security Review

06 Playbooks/Security Review.md

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.

Knowledge connections