🛡🛡🛡 Anthropic has launched Project Glasswing - an initiative to protect critical software using the new Claude Mythos Preview model.
Mythos Preview is an unreleased frontier model designed to find vulnerabilities. On CyberGym (vulnerability reproduction), it scores 83.1% compared to 66.6% for Opus 4.6. On SWE-bench Verified, it scores 93.9% against 80.8%. Essentially, the model finds software flaws at the level of top human specialists.
The specific findings are impressive: a 27-year-old vulnerability in OpenBSD allowing remote crashes, a 16-year-old bug in FFmpeg missed by automated tools despite 5 million test runs, and a chain of vulnerabilities in the Linux kernel for privilege escalation.
The coalition includes AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic is allocating $100M in credits for model use and $4M in donations to open-source security organizations. Over 40 organizations managing critical infrastructure will gain access.
Dario Amodei wrote in his post that the model "should feel terrifying" and that he is proud of the decision to give it first to cyber defenders rather than releasing it to the general public.
It's frightening to realize that all this power is just the beginning.
