Project Glasswing at Scale: 150 Organisations, 10,000 Critical Bugs, and What It Means for Software Security
Anthropic has expanded Project Glasswing from its initial 11 partner organisations to approximately 150 organisations across 15 or more countries. The combined output of this coalition is striking: more than 10,000 HIGH or CRITICAL severity software vulnerabilities found and disclosed. To put that in context, a typical year of public CVE disclosure produces somewhere in the range of 25,000 to 30,000 total CVEs across all severity levels. Glasswing's coalition found 10,000 high-severity issues in five weeks.
This is not incremental improvement in security tooling. This is a structural change in the rate at which critical vulnerabilities can be discovered and disclosed at scale.
The Coalition That Makes This Credible
Project Glasswing's expanded coalition includes organisations that collectively own or maintain enormous amounts of critical software infrastructure:
- Cloud infrastructure: AWS, Microsoft, Google
- Hardware and networking: NVIDIA, Cisco, Broadcom
- Financial services: JPMorgan Chase
- Security specialists: Palo Alto Networks, CrowdStrike
- Consumer platforms: Apple
- Open-source foundations: The Linux Foundation
When organisations of this scale are running Claude Mythos against their own codebases and critical dependencies, the surface area being scanned is an enormous fraction of the software the global economy runs on.
What the 10,000 Number Actually Means
Ten thousand high and critical severity vulnerabilities in five weeks represents roughly a year of normal CVE disclosure pace, compressed into a single coalition effort. The implications are significant:
- Many of these vulnerabilities exist in open-source libraries that underpin applications across industries — every organisation consuming those libraries is exposed until patches ship
- The coordinated disclosure model means fixes are being developed alongside the discovery, rather than vulnerabilities sitting undisclosed for months
- The geopolitical dimension matters: 15 countries means cross-border vulnerability coordination at a scale that has no clear precedent in the security industry
What Every Developer Consuming Open-Source Should Do Now
If your CI/CD pipeline does not already include automated dependency scanning with severity gating, that needs to change. Glasswing's disclosures are flowing into the NVD and vendor advisories. Your dependency update cadence needs to match the accelerated pace of disclosure that AI-powered scanning is producing.
- Enable automated dependency scanning in your CI pipeline — GitHub Dependabot, Snyk, or Trivy are all viable starting points
- Set severity gating at HIGH and above for production deployments — do not ship with known high-severity unpatched dependencies
- Audit your direct and transitive dependency surface — the vulnerabilities being found are often in widely-used transitive dependencies, not just direct imports
- Subscribe to security advisories for your critical dependencies — passive discovery of patches is too slow in the current environment
Honest Trade-offs in the Glasswing Model
- Closed access: Claude Mythos is not publicly available — the scanning capability is gated behind partnership agreements, which means smaller organisations cannot directly participate
- Vendor concentration: The most powerful security scanning capability in the industry is controlled by a single AI lab — a supply chain risk in itself
- Disclosure coordination complexity: Cross-border coordinated disclosure at this scale creates legal and jurisdictional complexity that the security community has not encountered before
- Patch lag: Finding 10,000 vulnerabilities is only valuable if the downstream ecosystem can absorb patches at the same rate — maintainer capacity remains a bottleneck
Key Takeaways
- Project Glasswing has expanded to 150 organisations and found over 10,000 high and critical severity vulnerabilities in five weeks — a structural acceleration of vulnerability discovery
- The coalition includes the organisations that own and maintain the software infrastructure the global economy depends on
- Every developer consuming open-source dependencies should treat this as a signal to tighten dependency scanning and patch cadence immediately
- The closed-access model and vendor concentration are legitimate trade-offs that enterprise security architects should acknowledge in their risk assessments
- OpenAI's Daybreak announcement in the same week signals that AI-powered security scanning is becoming a contested market, not an Anthropic monopoly
