Project Glasswing and Mythos Preview — The Full Picture of All 12 Founding Partners
This article provides a comprehensive overview of Project Glasswing and its dedicated model, Claude Mythos Preview, announced by Anthropic in April 2026. We cover the 12 founding partners, benchmark results that surpass Opus 4.6 (including CyberGym at 83.1%), up to $100 million in credit support, and the conditions under which users and organizations in Japan can gain access — all based on official primary sources.
Project Glasswing is a cross-industry cybersecurity initiative announced by Anthropic on April 7, 2026, bringing together 12 founding partners including AWS, Apple, Google, Microsoft, and NVIDIA, along with more than 40 organizations responsible for critical infrastructure, to build a global framework that establishes a defensive advantage.
The dedicated model Claude Mythos Preview dominates Opus 4.6 with CyberGym at 83.1% and SWE-bench Verified at 93.9%, and has demonstrably discovered vulnerabilities that went undetected for 27 years in OpenBSD and 16 years in FFmpeg. Participating organizations receive up to $100 million in usage credits, and a combined $4 million in OSS donations is being made to Alpha-Omega and Apache.
Mythos Preview has no planned general release, and access is limited to founding partners, certified organizations, OSS maintainers, and Cyber Verification Program applicants. Anthropic has indicated a future plan to gradually integrate Mythos-class capabilities into the next Opus model, and there is currently no direct means for individual developers to access it.
目次 (13)
- What Is Project Glasswing — Announced 2026-04-07, a Cross-Industry Initiative to Protect Critical Infrastructure with AI
- The Full List of 12 Founding Partners — AWS / Apple / Google / Microsoft / NVIDIA / Anthropic and More
- Claude Mythos Preview — A Limited Frontier Model Dedicated to Glasswing
- Proven Vulnerability Detection — 27 Years in OpenBSD, 16 Years in FFmpeg, Linux Kernel Privilege Escalation
- Benchmark Comparison — CyberGym 83.1% / SWE-bench Verified 93.9%, Dominating Opus 4.6 Across the Board
- Intended Use Cases — Limited to Cybersecurity, No Mention of General Business Use
- Positioning vs. Existing Claude Products — Mythos Preview Is Limited-Release with No Plans for General Availability
- Investment and Support — Up to $100 Million in Credits, $4 Million in OSS Donations
- Perspective for Users in Japan — Who Has Access and the Current State of Japanese Support
- Who Can Access It Now — Founding Partners / Certified Organizations / OSS Maintainers / Cyber Verification
- Japanese Language Support and Japan Market Information — Not Mentioned on Official Page, Awaiting Updates
- Addendum as of June 2026
- Sources (Primary Information)
What Is Project Glasswing — Announced 2026-04-07, a Cross-Industry Initiative to Protect Critical Infrastructure with AI
Project Glasswing is a cross-industry cybersecurity initiative announced by Anthropic on April 7, 2026 source.
Anthropic determined that it was urgent to establish a defensive advantage before AI models' vulnerability-detection capabilities could proliferate to malicious actors. Working with major industry players, the initiative builds a framework to protect critical global software infrastructure using AI source.
The Full List of 12 Founding Partners — AWS / Apple / Google / Microsoft / NVIDIA / Anthropic and More
Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks source.
In addition to these 12 founding partners, more than 40 organizations that build and maintain critical software infrastructure are eligible to participate source.
Claude Mythos Preview — A Limited Frontier Model Dedicated to Glasswing
Claude Mythos Preview is a limited-release frontier model developed specifically for Project Glasswing. Anthropic describes it as having "coding capabilities that surpass everyone except the very highest-level human security experts" source.
Proven Vulnerability Detection — 27 Years in OpenBSD, 16 Years in FFmpeg, Linux Kernel Privilege Escalation
- OpenBSD: Discovered a vulnerability that had gone undetected for 27 years
- FFmpeg: Identified a bug that had gone undetected for 16 years (in a location tested 5 million times by automated tools)
- Linux Kernel: Automatically generated a privilege escalation attack by chaining multiple vulnerabilities
Source: Anthropic: Project Glasswing
Benchmark Comparison — CyberGym 83.1% / SWE-bench Verified 93.9%, Dominating Opus 4.6 Across the Board
Comparing key benchmarks between Mythos Preview and Opus 4.6, Mythos leads in every category.
| Benchmark | Mythos Preview | Opus 4.6 |
|---|---|---|
| CyberGym (vulnerability reproduction) | 83.1% | 66.6% |
| SWE-bench Verified | 93.9% | 80.8% |
| SWE-bench Pro | 77.8% | 53.4% |
| Terminal-Bench 2.0 | 82.0% | 65.4% |
| SWE-bench Multilingual | 87.3% | 77.8% |
| GPQA Diamond (with tools) | 64.7% | 53.1% |
Notes on each metric: SWE-bench Verified = how well the model can independently resolve real GitHub issues / SWE-bench Pro = resolution rate on higher-difficulty real-world tasks / Terminal-Bench 2.0 = task completion rate for terminal operations / SWE-bench Multilingual = resolution rate on multi-language repositories beyond Python / GPQA Diamond = accuracy on expert-level science and technology questions (with tools)
Source: Anthropic: Project Glasswing
Intended Use Cases — Limited to Cybersecurity, No Mention of General Business Use
The intended use cases listed on the official page are limited to the cybersecurity domain, with no mention of general business applications source. The officially listed use cases are as follows:
- Vulnerability scanning and remediation for critical infrastructure software
- Security audits of open-source projects
- Defensive threat research
Positioning vs. Existing Claude Products — Mythos Preview Is Limited-Release with No Plans for General Availability
Claude Mythos Preview is clearly distinguished from the publicly available model lineup offered through Claude.ai and the Claude API.
| Product | Audience | Availability |
|---|---|---|
| Claude.ai | General users | Public |
| Claude API (Opus 4.7, etc.) | Developers | Public |
| Claude Code | Software developers | Public |
| Claude Mythos Preview | Glasswing participants and certified organizations | Limited |
Anthropic has explicitly stated that "there are no plans to make Mythos Preview publicly available," while indicating that Mythos-class capabilities will be gradually integrated into future Opus models source.
Investment and Support — Up to $100 Million in Credits, $4 Million in OSS Donations
Participating organizations receive up to $100 million in credits, and OSS organization donations total $4 million.
| Support | Amount |
|---|---|
| Mythos Preview usage credits (for participating organizations) | Up to $100 million (USD) |
| Donations to OSS security organizations (total) | $4 million (USD) |
| ├ Alpha-Omega (OpenSSF) | $2.5 million (USD) |
| └ Apache Software Foundation | $1.5 million (USD) |
Source: Anthropic: Project Glasswing
Perspective for Users in Japan — Who Has Access and the Current State of Japanese Support
Who Can Access It Now — Founding Partners / Certified Organizations / OSS Maintainers / Cyber Verification
Access to Claude Mythos Preview is limited to the following source:
- Project Glasswing founding partners (including access via AWS, Google, Microsoft, and other cloud platforms)
- More than 40 certified critical infrastructure organizations
- OSS maintainers (can apply via the Claude for Open Source program)
- Applicants to the Cyber Verification Program (in preparation)
Japanese Language Support and Japan Market Information — Not Mentioned on Official Page, Awaiting Updates
The official Glasswing page contains no information about Japanese language support, Japan market localization, or participation by Japanese companies source. Details on pricing, Japanese-language support, and domestic partners have not been announced at this time.
We will update this article as further information becomes available.
Addendum as of June 2026
This article was written based on information available at the time of the April 2026 announcement. Here is a supplemental update on developments since then.
As successors to Opus 4.6, which is used as a comparison point in this article, Opus 4.7 and Opus 4.8 (supporting up to 1 million tokens of context) have been made publicly available. You can find details in our article Claude Opus 4.8 | 1M Context and Next-Generation Reasoning Capabilities.
Anthropic had indicated plans to gradually integrate Mythos-class capabilities into future Opus models, but the completion of that integration has not been officially confirmed as of June 2026. We recommend avoiding definitive conclusions and checking Anthropic's official announcements directly.
The access conditions for Claude Mythos Preview (limited to founding partners, certified organizations, and OSS maintainers) remain unchanged from the original announcement.
Sources (Primary Information)
The following primary sources were directly referenced in the creation of this article. Please always check the links below for the latest and most accurate information.
- Anthropic: Project Glasswing — Official page (announced 2026-04-07, accessed 2026-04-23)
- Anthropic: News — News index (accessed 2026-04-23)
