Project Glasswing: How a Secret Claude Model Found 23,000 Software Bugs
Anthropic announced Project Glasswing — an industry coalition that used an unreleased Claude model to scan over 1,000 open-source software projects for security vulnerabilities. The results are striking: 23,019 issues found, with 6,202 classified as high or critical severity. Cloudflare identified 2,000 bugs in their own codebase. Mozilla fixed 271 Firefox vulnerabilities — ten times the volume the prior AI tool found. This is the most significant AI-driven security initiative announced so far.
Who Is In the Coalition
The partner list reads like a who's who of enterprise technology: AWS, Apple, Google, Microsoft, JPMorgan, Cisco, NVIDIA, Palo Alto Networks, CrowdStrike, Broadcom, and the Linux Foundation. The breadth of that coalition is itself a signal — this is not an Anthropic marketing exercise. Organisations with serious security postures and legal exposure signed on because the results warranted it.
Anthropic committed 100 million USD in model usage credits to the initiative plus 4 million USD in direct donations to open-source security projects. These are material commitments, not token gestures.
What Claude Mythos Preview Can Do That Public Claude Cannot
The model used in Glasswing is Claude Mythos Preview — an unreleased variant not available through the public API. Based on what Anthropic has shared, Mythos Preview has extended capabilities specifically tuned for:
- Deep codebase traversal: Analysing inter-file dependencies and call graphs at a scale that general Claude models do not consistently handle
- Vulnerability pattern recognition: Trained specifically on CVE databases, secure code guidelines, and known exploit patterns across languages
- Low false positive rate: Security scanning is only useful if the signal-to-noise ratio is high enough that human reviewers trust the output — Mythos Preview was tuned for precision over recall
Anthropic is keeping Mythos closed because the same capability that finds vulnerabilities can describe how to exploit them. This dual-use concern is why it is running in a controlled coalition rather than as a public API.
The 23,000 Vulnerability Breakdown in Perspective
23,019 issues across 1,000-plus projects sounds large — and it is — but context matters:
- 6,202 high and critical issues represent about 27 percent of total findings — a meaningful ratio suggesting the model was not flagging trivial style issues as security problems
- Mozilla's 271 Firefox vulnerabilities represent a 10x improvement over the prior AI tool they used — the baseline comparison matters as much as the absolute number
- Cloudflare's 2,000 findings came from a codebase maintained by a team with sophisticated security practices — this was not low-hanging fruit
What This Means for Your Security Posture
The implications for enterprise developers and security architects are practical:
- Your open-source dependencies carry risk you have not quantified: The projects Glasswing scanned are the same projects inside your supply chain. If 23,000 issues were found, some percentage exist in code you are running in production.
- AI-driven SAST is becoming production-grade: Static application security testing has historically had noise problems that limited adoption. Glasswing results suggest AI models are reaching precision thresholds where the signal is worth the review cost.
- The gap between AI security capability and your current tooling is widening: Teams still running basic linters and pattern-match SAST are operating with a fundamentally different capability ceiling than what Glasswing demonstrated.
3 Moves to Make This Week
- Audit your critical open-source dependencies: Check whether the projects you depend on were included in Glasswing's scope and whether patches were released. The Linux Foundation will publish a disclosure schedule.
- Add AI-assisted code review to your CI pipeline: You do not have Mythos Preview, but public Claude, GPT-5.5, and specialised tools like GitHub Advanced Security with Copilot provide meaningful security scanning today. Start with your highest-risk repositories.
- Apply for Glasswing access: Anthropic has indicated they will extend the programme. If you maintain open-source software or run a security team at a qualifying organisation, reach out through Anthropic's enterprise security channel.
Key Takeaways
- Project Glasswing used an unreleased Claude model to find 23,019 vulnerabilities across 1,000-plus open-source projects — including 6,202 high and critical severity issues
- The coalition includes AWS, Apple, Google, Microsoft, JPMorgan, and Cisco — organisations with serious security postures who validated the results before signing on
- Claude Mythos Preview is closed because its dual-use capability requires controlled access — the same power that finds vulnerabilities can describe exploits
- The practical implication: your open-source supply chain has unquantified risk that AI-driven scanning can now surface at scale
- AI-assisted security scanning is moving from experimental to production-grade — teams not investing in this capability are falling behind on security posture
