Project Glasswing
Sources: Anthropic — Project Glasswing · Project Glasswing announcement paper page
Project Glasswing is a coalition initiative announced by Anthropic on May 12, 2026 to apply Claude Mythos Preview (an unreleased frontier model) to defensive vulnerability discovery on the world’s most-critical software. Twelve named partners and 40+ additional organizations participate; Anthropic has committed up to $100M in usage credits plus $4M in direct donations to open-source security organizations. The initiative is framed as a 90-day-report cadence preview, with explicit national-security positioning and a longer-term goal of seeding an independent third-party body to coordinate large-scale AI-augmented defensive cybersecurity.
One-Month Update (2026-05-22)
The first progress report (Glasswing initial update) records that the ~50 partners have collectively found more than ten thousand high- or critical-severity vulnerabilities in one month, with several reporting bug-finding rates up more than tenfold. The headline reframing: discovery is no longer the constraint; verification, disclosure, and patching are.
Named partner and evaluator results:
| Organization | Result |
|---|---|
| Cloudflare | 2,000 bugs (400 high/critical) across critical-path systems; FP rate better than human testers |
| Mozilla | 271 vulnerabilities found and fixed in Firefox 150, >10× the Firefox 148 count under Opus 4.6 |
| UK AI Security Institute | Mythos first model to solve both AISI cyber ranges end to end |
| Palo Alto Networks | Latest release shipped >5× the usual number of patches |
| Oracle | Finding and fixing vulnerabilities multiple times faster |
| One partner bank | Mythos helped prevent a fraudulent $1.5M wire transfer |
In parallel, Anthropic’s open-source scanning program (1,000+ projects) estimated 6,202 high/critical vulnerabilities, of which 1,752 have been assessed (90.6% true positives) and only 75 patched so far, the funnel that gives the bottleneck-inversion its empirical shape. Maintainer overload is a named constraint: some maintainers asked Anthropic to slow down disclosures. See the update page for the full funnel and tooling releases (Claude Security public beta, the Cyber Verification Program, shared skills/harness/threat-model-builder).
Coalition Partners
Named launch partners (12)
| Partner | Wiki page | Role / Quote attribution |
|---|---|---|
| Amazon Web Services | AWS | Amy Herzog (VP & CISO): testing Mythos in AWS security operations |
| Anthropic | Anthropic | Model vendor; initiative lead |
| Apple | (no wiki page yet) | (no public quote) |
| Broadcom | (no wiki page yet) | (no public quote) |
| Cisco | (no wiki page yet) | Anthony Grieco (SVP & CSTO): “AI capabilities have crossed a threshold” |
| CrowdStrike | CrowdStrike | Elia Zaitsev (CTO): “the window … has collapsed” |
| Heather Adkins (VP Security Engineering): Mythos via Vertex AI | ||
| JPMorganChase | (no wiki page yet) | Pat Opet (CISO): financial-system framing |
| The Linux Foundation | (no wiki page yet) | Jim Zemlin (CEO): OSS maintainer access |
| Microsoft | Microsoft | Igor Tsyganskiy (EVP Cybersecurity + Microsoft Research): CTI-REALM evaluation |
| NVIDIA | NVIDIA | (no public quote) |
| Palo Alto Networks | Palo Alto Networks | Lee Klarich (CPTO): “more attacks, faster attacks, more sophisticated attacks” |
Additional organizations
40+ further organizations have access to Mythos Preview to “scan and secure both first-party and open-source systems.” The post does not enumerate them; the 90-day public report may.
Glasswing’s method has one documented instance of reaching outside the coalition’s published membership. Visa’s visa-vulnerability-agentic-harness is described in Semgrep’s July 2026 survey as an agentic static-analysis pipeline built on learnings from Project Glasswing, released under Apache 2.0 and closed to external contributions.1 The description is Semgrep’s LLM-generated repository summary rather than a statement by Visa or Anthropic, and it places Visa nowhere in relation to the 40+ organizations holding Mythos Preview access. The instance documents the method’s reach into third-party open source rather than Visa’s standing in the coalition. See VVAH for the harness itself.
Partner terms
- Access to Claude Mythos Preview during the research preview period.
- Coverage of substantial usage via Anthropic’s $100M usage-credit commitment.
- Post-preview pricing: $25 / $125 per million input/output tokens (Glasswing-participant rates). Available on Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
- Information-sharing: partners “to the extent they’re able” share information and best practices with each other.
- Public reporting by Anthropic within 90 days on lessons learned, vulnerabilities fixed, and improvements that can be disclosed.
Operational Focus Areas
Per the announcement, Glasswing partner work concentrates on:
- Local vulnerability detection: finding bugs in partner-owned code.
- Black-box testing of binaries: pentest-style coverage where source isn’t accessible.
- Securing endpoints: defender-side product enhancement.
- Penetration testing of systems: internal adversarial use of Mythos.
Industry-Standards Component
Anthropic commits to “collaborate with leading security organizations to produce a set of practical recommendations for how security practices should evolve in the AI era.” Named candidate areas:
- Vulnerability disclosure processes
- Software update processes
- Open-source and supply-chain security
- Software development lifecycle and secure-by-design practices
- Standards for regulated industries
- Triage scaling and automation
- Patching automation
Donations
Direct cash donations to open-source security organizations:
- $2.5M to Alpha-Omega and OpenSSF (via the Linux Foundation).
- $1.5M to the Apache Software Foundation.
- Claude for Open Source program: additional access for OSS maintainers via claude.com/contact-sales/claude-for-oss.
National-Security Positioning
Anthropic frames Glasswing as a defensive imperative against state-sponsored threats (named: China, Iran, North Korea, Russia). Explicit language: “The US and its allies must maintain a decisive lead in AI technology.” Anthropic discloses ongoing discussions with US government officials about Mythos’s offensive and defensive cyber capabilities. The long-term proposed structure is an independent third-party body bringing private- and public-sector organizations together.
Wiki Position
Glasswing is the organizing artifact for the wiki’s ai-in-sec-defense axis as of May 13, 2026, and for Frontier AI for Vulnerability Discovery within it. It supersedes the wiki’s previous treatment of defender-AI as a vendor-by-vendor productized capability with a coalition-backed industrial-scale framing. Semgrep’s July 2026 market-structure finding sits at a different layer and does not contradict this: no reference open-source harness will emerge today, and many organizations will build their own “shop jigs” for vulnerability finding, so where Glasswing consolidates model access, the harness layer built around that access is diffusing rather than consolidating.1
Critical context:
- Glasswing is not a product; it is a coalition initiative. The product (model) is Claude Mythos Preview.
- Glasswing partners include both AI vendors and AI-adopting enterprises: the model vendor (Anthropic), other AI vendors (Google, Microsoft), CSPs (AWS), security vendors (CrowdStrike, Palo Alto Networks), infrastructure providers (NVIDIA, Cisco, Broadcom, Apple), foundations (Linux Foundation), and financial institutions (JPMorganChase).
- The Linux Foundation’s inclusion is structurally important: it signals OSS maintainer reach beyond commercial product customers.
- Microsoft’s MDASH is a Glasswing artifact: Microsoft’s “generally available AI models” silence in the MDASH announcement is explained by coordinated-launch constraints. Mythos is almost certainly one of MDASH’s orchestrated models.
Open Questions
- The 40+ additional organizations beyond the named 12.
- Specifics of OSS maintainer access via Claude for Open Source.
- Per-partner case studies: the 90-day public report should surface some.
- Government / standards-body engagement beyond the high-level mention.
- Whether the proposed “independent third-party body” materializes; what its charter would be.
- How responsible-disclosure timelines align across 50+ partners using the same model on the same codebases.
See Also
- Project Glasswing announcement paper page: source summary.
- Claude Mythos Preview: the model.
- Anthropic: initiative lead.
- XBOW’s Mythos Evaluation: independent (non-partner) third-party evaluation of Mythos.
- MDASH announcement: Glasswing-partner defender-AI artifact.
- CyberGym: benchmark on which both Mythos (raw, 83.1%) and MDASH (88.45%) sit.
- Frontier AI for Vulnerability Discovery: wiki thesis Glasswing anchors.
- SDLC in the AI-Attacker Era: adjacent thesis directly supported by Glasswing’s framing.
Notes
Footnotes
-
Semgrep — Comparing open source AI code security harnesses, July 2026 (no day-level date exposed; author not named). The market-structure argument (no reference open-source harness, “shop jigs”) is human-written; the VVAH description (“built on learnings from Project Glasswing,” licence, closed-to-contributions) is from Semgrep’s LLM-generated repository summary. Summarized at OSS AI Security Harness Comparison. ↩ ↩2