AGENT SKILL

security-review

Review a change for risks to trust boundaries, input handling, access control, secrets, sensitive data or dependencies. Use when a change affects input, authorization, data exposure, security configuration or dependencies; use threat-modeling for a design before it is built.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 1
Forks
0

01 前言 Front Matter

name
security-review
description
Review a change for risks to trust boundaries, input handling, access control, secrets, sensitive data or dependencies. Use when a change affects input, authorization, data exposure, security configuration or dependencies; use threat-modeling for a design before it is built.

02 资源文件 1

03 SKILL.md 正文

Security review

When to use

In the review phase of any change that touches external input, access control, persistence of sensitive data, logging, configuration, dependencies or infrastructure. When the change adds a new component, interface or data flow, run threat-modeling first: a review checks the code against threats someone has named.

Required context

None is required. Useful: the integrations and architecture context that identify trust boundaries (what is external, who calls what), and project security rules.

Preconditions

The diff is final enough to review.

Procedure

  1. List the trust boundaries the change touches: where data enters from users, other services, files, queues or configuration.
  2. Check each entry point: input is validated at the boundary, and output is encoded for its destination (queries, markup, commands, file paths, logs).
  3. Check authentication and authorization on every new or changed operation. A missing check is a finding even if "the interface does not expose it".
  4. Check secrets and sensitive data: nothing written to code, logs, errors, evidence or test data; nothing returned to a caller who should not see it.
  5. Check new dependencies: why they are needed, where they come from, whether they are maintained, and whether they bring known vulnerabilities.
  6. Check failure behavior: errors do not leak internals, and a failed security check denies rather than allows.
  7. Run the security checks the profile has.
  8. Trace each candidate from source to sink before calling it a finding: name the untrusted source, the path it takes through the changed code and the sink it reaches, and confirm no guard on that path already stops it. A candidate with no nameable source, or one stopped by an existing guard, is dropped; one the diff did not introduce is raised separately, not recorded here. When the agent can, have a separate pass re-trace each finding from the code.
  9. Record each finding so its owner can act on it in one read: state what the code currently does, the concrete impact, the suggested correction, and the consequence of leaving it as-is. Include severity, the boundary and traced path. Keep hardening ideas without a traced path in notes. Hardening ideas that no traced path demands are notes, kept apart from findings. Fix what is in scope; raise what is not. With no finding, list the boundaries examined and why each holds.

Tools

core.repository.diff to review exactly what changed.

Rules

core.security.no-secrets-in-source always; project security rules by scope.

Verification

Required: security — a security check when the profile has one, otherwise a gap stating that only the manual review below was done. Recommended: static-analysis.

Evidence

A review artifact listing the boundaries examined and the findings. An agent-recorded review supports no claim by itself; it tells the human reviewer where to look.

Completion criteria

  • Every touched boundary was examined and is listed in the review.
  • Every finding carries a traced source-to-sink path.
  • No unresolved finding of high severity remains in scope.

Failure modes

Failure Signal Response
Review of intent, not code Findings describe what the code should do Re-read the diff
Trusting internal callers "Only called internally" with no enforcement Treat as a boundary or record the assumption
Scanner as the review Only tool output is reported Walk steps 1 to 6 as well
Theoretical finding No source, path or sink can be named for it Trace it (step 8) or turn it into a note

References

None.

04 同仓库更多 Skills