TL;DR

Anthropic has restricted its Claude Mythos Preview model to a vetted group of technology giants and cybersecurity firms, saying the system is too capable at finding software vulnerabilities to release publicly. Mythos has already uncovered thousands of unpatched flaws, including a 27-year-old bug in OpenBSD and defects in widely used browsers and operating systems.

A deliberate hold-back

Mike Krieger of Anthropic Labs told the HumanX conference the model would not be released to the public. Instead, it is being shared under Project Glasswing with Amazon, Apple, Microsoft, Cisco, Broadcom, CrowdStrike, Palo Alto Networks and the Linux Foundation — roughly 40 organisations in total, backed by around $100 million in Anthropic compute. The company says engineers with no security training have asked Mythos to find remote code execution bugs overnight and woken up to complete, working exploits.

Why it matters for UK defenders

The move lands as the NCSC and UK critical infrastructure operators are already warning about AI-assisted vulnerability discovery compressing patch windows from months to minutes. CrowdStrike’s CTO Elia Zaitsev called the window between discovery and exploitation “collapsed”. For UK businesses running on widely deployed operating systems and browsers, Glasswing means defenders at a handful of large vendors get early sight of flaws their own security teams cannot yet see — a centralisation of defensive advantage that mirrors, in reverse, concerns UK regulators have raised about concentration of offensive capability.

This is also the first time Anthropic has limited a general-purpose model release on cyber grounds, a line the company foreshadowed in its responsible scaling policy but had not previously enforced. It arrives days after an internal leak exposed Mythos documentation, and a week after Claude Code source code was made public — incidents Anthropic blamed on human error and which prompted fresh questions about its own security posture.

Looking forward

Glasswing will not stay unique for long. If Mythos-class capability emerges at a frontier lab with weaker release controls, the defenders’ head start disappears. UK CISOs should assume adversary access to similar tooling within 12 months and audit legacy code accordingly.