Anthropic Unveils Project Glasswing with Claude Mythos Preview, Keeps New Model Out of Public Hands

Anthropic has just announced Project Glasswing, and it's far from your standard product launch. It is a new AI model as part of cybersecurity initiative backed by Apple, AWS, Broadcom, Cisco, CrowdStrike, Google, Microsoft, NVIDIA, and Palo Alto Networks to proactively hunt down and patch vulnerabilities in the world's most critical software infrastructure.

At the center of the announcement is Claude Mythos Preview, a general-purpose frontier model that Anthropic has deliberately kept unreleased to the public. According to the company, Mythos Preview has autonomously identified thousands of high-severity, zero-day vulnerabilities across all major operating systems and web browsers.

Some of these flaws have survived decades of human review and millions of automated tests, highlighting just how capable AI has become at reading and exploiting code. For example, Mythos Preview unearthed a 27-year-old remote crash bug in OpenBSD, a famously hardened OS. It also found a 16-year-old flaw in FFmpeg missed by automated testing five million times. And it chained multiple Linux kernel exploits to achieve total system control.

The model is privately available to Glasswing partners, which include JPMorgan Chase, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Palo Alto Networks, and about 40 other critical infrastructure organizations. Anthropic is also subsidizing usage with $100 million in credits and donating $4 million to open-source security groups.

According to The Verge, the company has briefed US government officials on Mythos Preview's capabilities and is "committed to working closely with all levels of government." The work ahead will take months, even years, and requires cooperation across AI developers, software maintainers, cyber defenders, and governments.