How to turn AI-driven vulnerability discovery into a faster, more trusted response

November 2026

The short version

  • AI now finds software flaws faster than most teams can triage them. Anthropic reports its partners found more than 10,000 high- or critical-severity vulnerabilities in a single month.
  • A finding is just a software name and a version. Your CMDB turns it into an answer: where you run it, who owns it, and what breaks if it is exploited.
  • The best teams will not argue with the AI. They will make its work auditable, verify it, and use every gap it exposes to fix the CMDB.

What changed

In April 2026, Anthropic launched Project Glasswing. About 50 partner organizations got access to an unreleased model, Claude Mythos Preview, to find and fix flaws in critical software before attackers could. Its May 22 update reported more than 10,000 high- or critical-severity vulnerabilities across partner software.

Finding flaws is no longer the bottleneck. Triage, reporting, and patching are. That work runs through your teams and your data, and configuration management sits at the center of it.

The pressure shows up in three ways: volume (which of our thousands of CIs matter most?), speed (disclosure windows are short, so a quarterly-accurate CMDB is not enough), and visibility (missing installs and broken relationships slow remediation, and leadership will see where).

Make the AI's work auditable

When an AI finding lands, the reflex is to challenge it. What did it query? Why did it flag this CI? Those are the right questions, but asking them to slow the work down leaves the exposure open. Answer them by design, before the first finding arrives.

Verification still matters. In Anthropic's open-source work, outside reviewers validated about 90.6% of the 1,752 critical-tier findings they triaged. That is strong, and it also means roughly one in ten did not hold up. Your process has to catch those.

Log every query, give the AI read-only access to only the tables it needs, require evidence (CI identifiers, relationship path, last-discovered date) with every conclusion, keep prompts and outputs on the record, and require a named human approver before any remediation task opens.

In a ServiceNow-based CMDB, start with the CI, application, database, software install, relationship, and hardware asset tables. Confirm table names and customizations in your own instance.

Four patterns to test

Leadership will soon ask: where are we exposed because our CMDB is wrong? Four patterns drive most of the answer.

PatternWhat to look forWhy it matters
Retired app, live infrastructureRetired applications still tied to operational servers or databases.Cost, plus unpatched systems left running.
Retired infrastructure, live appProduction applications tied to retired servers or databases.Risk. The record is wrong, or a service depends on something gone.
Missing relationshipsProduction apps with no infrastructure, or servers with no application.Blind spots. Impact and ownership become guesses.
Stale or conflicting recordsCIs not discovered recently, with no owner, or with conflicting status.Trust. Leaders cannot act on numbers they cannot defend.

How to run it

  1. Define terms. Agree on what counts as production, retired, and too old for discovery data.
  2. Test the four patterns. Claude queries read-only and returns the evidence behind every result.
  3. Verify. Check against a second source such as discovery, cloud inventories, or billing. A person reviews a sample and reports how many held up.
  4. Brief and fix. Give leadership one view by business service: exposure, cost, and confidence. Open owner-assigned tasks for every correction.

The real opportunity

Every finding Claude cannot map to a CI is a defect report on your CMDB, delivered faster than any audit cycle could. CMDB health stops being a hygiene project and becomes a measurable driver of how fast you close risk. Track it with a few measures:

  • Percentage of findings mapped to a CI with a known owner
  • Time from finding to a prioritized remediation task
  • Share of critical CIs with discovery data refreshed in the last 30 days
  • Findings the AI mapped that human review later overturned

Your next 30 days

  • Week 1. Name an executive owner across configuration management, security, and ITSM.
  • Week 2. Set up a read-only service account and logging.
  • Week 3. Pilot on one past vulnerability and compare with what your team did by hand.
  • Week 4. Turn the gaps into a funded CMDB improvement plan.

Notes: Glasswing figures come from Anthropic's May 22, 2026 update (anthropic.com/research/glasswing-initial-update); check its later public accounting for current numbers. Claude Mythos Preview is not publicly available. The controls above are design recommendations, not built-in features of any specific deployment. Table names and status values follow common ServiceNow conventions, so verify them in your instance.