Vulnerability triage that isn't repeated
DevOps, SRE and securityTeam triaging dependency advisories from a scanner
Our scanner flagged a high-severity advisory in an XML parser deep in the tree. An agent spent 30k tokens proving we never parse untrusted XML with it, so it isn't reachable. The next week the same advisory came up and a different agent did the same proof.
How do I stop AI agents from re-triaging the same dependency advisory every week?
Have the agent save its triage with ArcRoom's complete_task: the advisory, the verdict and why, tied to the git hash of the lockfile and the files it checked. Next run, find_task returns that verdict as a reuse match while the lockfile is unchanged. When the lockfile or those files change, the verdict goes stale and the agent triages again. The scanner's ignore rule still goes in its own config.
Also asked as: Our agent keeps proving the same vulnerability isn't reachable. How do we save that work? · How do I record why a scanner advisory doesn't affect us? · Can an AI agent reuse an earlier reachability check for a security advisory?
The agent saves the triage with complete_task: advisory, verdict, and why.
- Reused answer
- Saved answer
- Freshness check
find_task → match: reuse (0.95)
Step by step
- The agent saves the triage with complete_task: advisory, verdict, and why.
- The answer is tied to the git hash of the lockfile and the files it checked.
- Next run, find_task returns the verdict as a reuse match while the lockfile is unchanged.
- When the lockfile or those files change, the verdict is stale and the agent triages again.
What the agent sees
find_task → match: reuse (0.95)
"XML parser advisory: not reachable. Only used by scripts/import.ts on our own build files, never on user input."
pnpm-lock.yaml unchanged since savedCompared with today's tools
Scanner ignore file with a reason (e.g. Snyk or Dependabot dismissals)
Dismiss the advisory in the scanner with a written reason, so it stops alerting.
- Setup
- 10 min
- Ongoing
- dismiss each advisory by hand, ~5 min
ArcRoom
Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.
- Setup
- 2 min (npx arcroom init)
- Ongoing
- None
The scanner's dismissal is the right way to silence the alert; ArcRoom keeps the agent's reachability proof and flags it when the code it relied on changes.
Where it stops
ArcRoom doesn't stop the scanner from alerting. The ignore rule still goes in the scanner's own config.
Try it in your project
Works with the agents your team already uses, and any agent over MCP or the HTTP API.