Project Glasswing Partners and Findings 2026: Who's In and What They Found

Project Glasswing is a controlled-access program. Partners receive access to Claude Mythos Preview -- Anthropic's most capable model for vulnerability research -- in exchange for a commitment to coordinated vulnerability disclosure. Mythos tests the partner's own software. Findings are embargoed until patches are available. Partners provide operational feedback that informs Mythos's continued development.

The initial launch on April 7, 2026 included approximately 50 organizations: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks on the commercial side. Open-source maintainers for Firefox (Mozilla), nginx, FreeBSD, OpenBSD, FFmpeg, and others. Plus Cloudflare, whose published findings have become the most detailed public case study of what Glasswing produces at scale.

The June 2, 2026 expansion added approximately 150 additional organizations and meaningfully broadened the scope. New sectors: power generation and distribution, water systems, healthcare infrastructure, communications networks, and semiconductor hardware. Geographic reach grew to 15+ countries. The expansion reflects a deliberate decision to stress-test Mythos against the software that runs critical infrastructure -- the category of target where successful exploitation has the highest societal consequence.

Cloudflare published the most detailed public accounting of a Glasswing deployment. Testing Mythos against their own infrastructure, they found 2,000 bugs -- 400 high/critical severity. False-positive rates were better than human testers, which is significant because false positives are the primary operational cost of automated security tooling. Cloudflare built an eight-stage, 50-parallel-agent harness to handle the throughput; single-agent approaches hit context limits before Mythos's capability was exhausted. Their assessment of Mythos's reasoning: it 'chains several small attack primitives together into a working exploit' in a way that resembles senior researcher work rather than pattern-matched vulnerability identification.

Mozilla tested Mythos against Firefox 150 and found 271 vulnerabilities compared to prior human-review findings in Firefox 148. The SpiderMonkey JavaScript engine results are the most technically significant: 14 of 18 proof-of-concepts generated by Mythos in a three-hour window were independently confirmed by Mozilla's own testing. SpiderMonkey is a high-value, frequently-audited target. The finding rate represents a substantial multiple over what existing security review processes were catching.

Beyond Cloudflare and Mozilla, most high/critical-finding partners reported more than a factor of ten increase in bug-finding rates compared to prior security review processes. Average patch time for high/critical Glasswing findings: two weeks, substantially faster than industry-average patch timelines for CVEs of comparable severity.

Anthropic committed $100M in Claude Mythos Preview usage credits to Glasswing partners -- the direct cost of running Mythos at scale across 200+ partner environments. The credit structure makes the program economically accessible to organizations, including open-source projects, that could not otherwise afford frontier AI-assisted security research.

Beyond partner credits, Anthropic made two targeted open-source investments: $2.5M to Alpha-Omega and the Open Source Security Foundation through the Linux Foundation, focused on the most critical open-source dependencies in the software supply chain. $1.5M to the Apache Software Foundation. These investments acknowledge that a significant fraction of the attack surface Glasswing is examining is open-source infrastructure that does not have a commercial security budget.

Claude Security, Anthropic's commercial product built on Claude Opus 4.7, entered public beta for organizations outside Glasswing. It patched 2,100+ vulnerabilities in its first three weeks. The commercial product serves two functions: it makes AI-assisted vulnerability research accessible to organizations that do not qualify for Glasswing, and it generates commercial revenue that funds continued Mythos development. The investment structure -- credits, open-source grants, and a commercial product path -- signals that Anthropic views AI-assisted vulnerability research as a long-term infrastructure commitment, not a research exercise with a defined end date.

Anthropic committed to a 90-day public summary report expected in July 2026. This will be the first major public accounting of aggregated Glasswing findings: fixed vulnerabilities, operational learnings, practical recommendations for how security practices should evolve given AI-scale vulnerability discovery. For practitioners building threat intelligence programs, the July report will be the most comprehensive single source on what Glasswing has actually found.

Following the report: high-volume patch releases are expected through H2 2026 as individual vulnerability embargo windows close. The CISA Known Exploited Vulnerabilities catalog will receive additions as Glasswing CVEs complete coordinated disclosure and enter the public record. Practitioners should build Glasswing monitoring into their intel workflows now rather than reacting to individual disclosures after the fact.

Practical monitoring approach: set alerts on red.anthropic.com for new disclosure publications, subscribe to NVD data feeds filtered for vendor/product combinations relevant to your environment, and watch the CISA KEV catalog for disclosure volume spikes in August and September 2026. The CVEs coming out of Glasswing will not all be labeled as such -- they will appear as standard CVEs from Mozilla, Cloudflare, AWS, and other partners. Cross-referencing publication dates and affected components against the Glasswing partner list is the most reliable way to identify them.