Skip to content

security.md

Latest commit

 

History

History
38 lines (32 loc) · 2.37 KB

security.md

File metadata and controls

38 lines (32 loc) · 2.37 KB
Raw

Security and vulnerability exposure

Use for trust boundaries, data, auth, dependencies, agent tools, CI, audits or uncertain/high risk. Keep checks within the authorized repository/environment.

Actions

  • Identify sensitive assets, entry points, trust boundaries and who can invoke the changed behavior. Trace access control and data flow through the affected path.
  • Review relevant input/output handling, injection, path traversal, object-level authorization, secret storage, unsafe deserialization, logging and data retention. For agents/MCP, include prompt injection, tool permissions and exfiltration paths.
  • Inspect dependency/lockfile and CI changes: version provenance, advisories, install scripts, credential scope, untrusted PR execution and artifact integrity. Use existing approved scanners where available; inspect how they handle data.
  • Triage scanner findings by affected version, reachability, exposure, confidence and impact. Verify current advisory data from authoritative sources when available. Scanner exit success is not a universal security claim.
  • Do not print secrets or upload proprietary code to an unapproved service. Record redacted evidence. If credentials are exposed, contain further disclosure and route rotation/revocation through the authorized owner and procedure.
  • Fix in-scope vulnerabilities and test the denied/abuse path. Track other findings with severity rationale, owner, remediation and any authorized exception expiry.

For a finding, distinguish vulnerable component presence from a reachable attack path and actual exposure. Record affected versions, principal/input, prerequisite, impact and uncertainty. Prioritize containment and remediation from that evidence. For agent/tool workflows, trace what a retrieved instruction could cause the tool to read, send or mutate; require existing authorization at that action boundary. After a fix, check both denied and allowed behavior without disabling the control.

Evidence and gate

security-review: affected boundaries reviewed, checks with tool/database date, findings disposition and residual uncertainty. Missing scanners or advisory access must be disclosed; source review may satisfy a source-review criterion, never an explicit required scan. Use private approved channels for sensitive findings; a framework invocation does not authorize public disclosure or active remote probing.