Security review
Use for any change touching sensitive areas. Details: security baseline, LLM application security.
Access
Section titled “Access”- Every new endpoint or page declares who may access it (deny by default)
- Per-object authorisation on the server
- No trust in client-supplied IDs, roles or prices
- Session and cookie settings unchanged or reviewed
Input and output
Section titled “Input and output”- Server-side schema validation for all new inputs
- Parameterised queries; no string-built SQL or shell
- Output encoded for context; no new raw-HTML escape hatches without sanitising
- Server fetches of user URLs allowlisted (SSRF)
- File uploads validated, size-limited, stored safely
Secrets and data
Section titled “Secrets and data”- No secrets in code, config, logs, prompts or client bundle
- Personal data minimised, tagged, and not logged
- New third-party processors added to the privacy list
Platform
Section titled “Platform”- Security headers and CSP still correct
- Rate limits on new expensive or sensitive endpoints
- New dependencies verified (exist, maintained, licence)
LLM features
Section titled “LLM features”- Untrusted content separated from instructions
- Model output schema-validated before use; never executed
- Tools least-privilege; risky actions need user confirmation
- Token, rate and cost limits
- Retrieval respects user permissions