Discussion

Study finds Claude Mythos access coincided with a surge in firms’ vulnerability disclosures

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#11431

A new analysis of Project Glasswing finds that firms given early access to Anthropic’s Claude Mythos Preview more than doubled their vulnerability disclosures compared with firms without access. The paper argues that this security boost may come with an uneven distribution of defensive capability: access was limited to a small group of large US technology companies.

Watch Desk analysis

What happened

Project Glasswing gave selected firms early access to the vulnerability-finding model. The paper’s author analysed public vulnerability data and reports that participating firms disclosed more vulnerabilities, found more high-impact flaws internally and received fewer reports from external bug-bounty researchers.

The analysis argues that frontier AI helped defenders find and fix flaws before exploitation, while the restricted rollout risks creating a security asymmetry between firms in different countries. The paper presents an analysis of public data, not proof that AI access alone caused the changes.

Why it matters

If the findings hold up, access to advanced AI security tools could affect who finds vulnerabilities first, and who has the resources to fix them. That makes the design of limited-access programmes a security question, not just a product-launch detail.

The reported fall in external bug-bounty reports also complicates the picture: more internal discoveries do not necessarily mean outside researchers have stopped contributing, but it does raise questions about how defensive work is measured and credited.

Our read

This is a useful early attempt to measure what frontier AI does in real-world defensive work, rather than settling for benchmark scores. The striking result is also a reminder that a tool can strengthen some defenders while leaving others behind. Treat the figures as evidence worth testing, not a final verdict on AI’s effect on cybersecurity.

What to watch

  • Whether follow-up research tests the results against other data or methods.
  • Whether access to comparable vulnerability-finding tools expands beyond the selected firms.
  • How future studies account for internal discoveries and external bug-bounty reports.

Discussion spark: Should access to powerful AI security tools be expanded broadly, even if that means accepting more risk, or kept limited to organisations best placed to use them safely?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.