Anthropic launched Project Glasswing in late April 2026, providing over 40 organizations with access to Claude Mythos Preview, an unreleased frontier model that has autonomously discovered thousands of high-severity zero-day vulnerabilities across major operating systems and web browsers. The initiative includes $100M in model usage credits and $4M in direct donations to open-source security organizations.
Partner Organizations Include Major Tech and Financial Institutions
Project Glasswing launch partners include AWS, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Additionally, over 40 organizations maintaining critical software infrastructure received access to the restricted model.
Anthropic committed $2.5M to Alpha-Omega and OpenSSF through the Linux Foundation, plus $1.5M to the Apache Software Foundation. The financial commitments aim to strengthen open-source security infrastructure before models with similar capabilities become broadly available.
Mythos Preview Discovered Decades-Old Vulnerabilities Autonomously
Claude Mythos Preview identified several critical vulnerabilities without human steering:
- OpenBSD: A 27-year-old vulnerability enabling remote system crashes
- FFmpeg: A 16-year-old flaw in video encoding code that automated tools had tested 5 million times without detection
- Linux kernel: Multiple chained vulnerabilities allowing user-to-administrator privilege escalation
- FreeBSD: A 17-year-old remote code execution vulnerability (CVE-2026-4747) allowing root access on machines running NFS
On the CyberGym vulnerability reproduction benchmark, Mythos Preview achieved 83.1% accuracy compared to Claude Opus 4.6's 66.6%. The model demonstrates superior performance across multiple coding tasks, including code generation and agentic problem-solving.
Model Remains Restricted Due to Offensive Capabilities
Mythos Preview remains unreleased to the general public because its powerful offensive capabilities could enable sophisticated cyberattacks if accessed by malicious actors. Anthropic stated that these capabilities emerged as a downstream consequence of general improvements in code, reasoning, and autonomy rather than through explicit training.
Anthropic plans to develop safeguards with a future Claude Opus model before broader deployment. The restricted release strategy allows refinement of defensive applications without exposing Mythos Preview's highest-risk outputs to potential attackers.
Controlled Release Prioritizes Critical Infrastructure Defense
By releasing Claude Mythos Preview initially to a limited group of critical industry partners and open-source developers, Anthropic aims to enable defenders to begin securing the most important systems before models with similar capabilities become broadly available. The strategy reflects growing concern about AI-enabled vulnerability discovery outpacing defensive capabilities.
The $100M in model usage credits provides participating organizations with substantial computational resources to conduct security assessments at scale. Combined with the $4M in direct donations, Project Glasswing represents one of the largest corporate commitments to open-source security infrastructure.
Key Takeaways
- Anthropic's Claude Mythos Preview autonomously discovered thousands of high-severity zero-day vulnerabilities across major operating systems and web browsers as part of Project Glasswing
- Notable discoveries include a 27-year-old OpenBSD vulnerability, a 16-year-old FFmpeg flaw tested 5 million times by automated tools without detection, and a 17-year-old FreeBSD remote code execution vulnerability (CVE-2026-4747)
- Mythos Preview achieved 83.1% accuracy on the CyberGym vulnerability reproduction benchmark compared to Claude Opus 4.6's 66.6%
- Anthropic committed $100M in model usage credits to over 40 partner organizations and $4M in direct donations to open-source security organizations including Alpha-Omega, OpenSSF, and Apache Software Foundation
- The model remains restricted from public release due to powerful offensive capabilities that could enable sophisticated cyberattacks, with safeguards planned for a future Claude Opus model before broader deployment