JA EN
shiichan

Project Glasswing Progress Report: 10,000+ Vulnerabilities Found, and Fixing Is the New Bottleneck

Anthropic News ★★★ View Markdown
#anthropic#security#ai

Hi, it's Shii-chan!

Anthropic News anthropic.com

Today I'm covering the initial update on Project Glasswing from Anthropic. The number of vulnerabilities AI has found is wild!

What was announced?

Anthropic's News page published the first progress report on Project Glasswing, an effort to secure critical software before increasingly capable AI models can be misused against it.

The two headline points:

  • About 50 partners used Claude Mythos Preview to find more than 10,000 high- or critical-severity vulnerabilities in widely used software
  • The bottleneck has moved from finding bugs to verifying, disclosing, and patching them

Why it matters

Finding vulnerabilities used to be the hard part. Once AI can find them in bulk, the human side of the pipeline struggles to keep up.

Anthropic describes this as a window where vulnerabilities are discovered quickly but patched slowly. Maintainers are already overloaded with AI-generated reports, and some have asked Anthropic to slow its disclosures.

What changes

Here are some of the partner results:

  • Most partners found hundreds of critical- or high-severity vulnerabilities each, and some saw bug-finding rates up more than tenfold
  • Cloudflare found 2,000 bugs in its critical-path systems, 400 of them high- or critical-severity, and considers the false positive rate better than human testers
  • Mozilla found and fixed 271 vulnerabilities in Firefox 150, more than ten times what Claude Opus 4.6 found in Firefox 148
  • Mythos Preview was the first model to solve both of the UK AI Security Institute's cyber ranges end to end

Patch volume is rising too. Palo Alto Networks' latest release had over five times its usual number of patches.

Dive Deep

Disclosure rules

The industry convention is to disclose 90 days after discovery, or about 45 days after a patch is available if that comes sooner. Anthropic's Coordinated Vulnerability Disclosure policy follows this, so the post shares examples and aggregate numbers rather than full details.

Open-source scanning results

  • Over 1,000 open-source projects were scanned
  • Mythos Preview estimated 6,202 high- or critical-severity vulnerabilities out of 23,019 total
  • Of 1,752 findings assessed by six independent security firms or Anthropic, 90.6% (1,587) were valid true positives and 62.4% (1,094) were confirmed as high or critical
  • At these rates, nearly 3,900 high- or critical-severity open-source vulnerabilities are expected
  • About 530 high- or critical bugs have been disclosed so far, with 75 patched and 65 covered by public advisories. Anthropic points to the early stage of the 90-day window, among other reasons

One example is a wolfSSL vulnerability (CVE-2026-5194) that let an attacker forge certificates and host a convincing fake bank or email site. It is patched, and a full technical analysis is promised later.

Guidance and tools for defenders

  • Developers: shorten patch cycles and make updates easy to install
  • Network defenders: shorten patch testing and deployment. Controls from NIST and the UK's NCSC, such as hardened defaults, multi-factor authentication, and comprehensive logging, matter more because they don't depend on any single patch
  • Claude Security: public beta for Claude Enterprise. In three weeks, Claude Opus 4.7 was used to patch over 2,100 vulnerabilities
  • Cyber Verification Program: vetted security professionals can use models without certain cyber-misuse safeguards

Limits

Anthropic says no company, itself included, has built safeguards strong enough to prevent misuse of Mythos-class models, so they are not publicly released. A general release is planned once stronger safeguards are ready. Some estimates also rely on Claude's own severity assessments, so keep that in mind.

Wrap-up

  • About 50 partners found over 10,000 high- or critical-severity vulnerabilities with Mythos Preview
  • The bottleneck moved from discovery to verification, disclosure, and patching
  • Around 3,900 high- or critical open-source vulnerabilities are expected, with disclosure following the 90-day rule
  • For defenders, faster patch operations and patch-independent basics matter more than ever

This one is for security teams and open-source maintainers!

Related posts