Knowledge base
CodexGuild Knowledge Base

CVE and dependency monitoring for agent-driven repos

as of Sep 22, 2026 · canonical · codexguild.com/kb/agent-cve-monitoring-playbook · exported 2026-10-11
Canonical as of Sep 22, 2026

CVE and dependency monitoring for agent-driven repos

Continuous (not incident-driven) monitoring: lockfiles committed, automated alerts, reachability triage, and agents that propose PRs with tests — not 3am bulk bumps.

CVE & dependency monitoring playbook

As of: 2026-09

Continuous monitoring (the boring 80%)

  • Commit lockfiles. Pin versions in apps (1.2.3, not ^1.2.3). Lockfile drift = CI failure.
  • Automated advisories: GitHub Dependabot / npm audit / pip-audit in CI, blocking on high severity.
  • Register manifests with CodexGuild (POST /v1/stack/manifests) → breaking-change + CVE alerts on your dashboard.
  • New dependency? Small PR with justification. Never bulk "chore(deps)" commits — especially not agent-made at 3am.

When an alert fires (the triage 20%)

  1. Reachability first (is the vulnerable API called in YOUR code?), severity second.
  2. Verify installed version from the lockfile, not memory.
  3. Read the fix's changelog for breaking changes before bumping.
  4. Agent PR = bump + regression tests + "why this is safe" note. Human merges.

Known-malicious ≠ known-vulnerable

A CVE gets you a patch window. A malicious package (TrapDoor, SANDWORM_MODE typosquats like claud-code, cloude-code) gets you incident response: isolate, rotate every credential the package's install-time environment could see, forensic the postinstall scripts it ran. OSSF malicious-package lists + install-time scanning catch these before execution — npm config set ignore-scripts with an allowlist is the nuclear option that works.