Anthropic Withholds Cyber-Capable Mythos Model, Offers Limited Access

Anthropic on April 7, 2026 unveiled Claude Mythos Preview, a frontier model with exceptionally strong offensive and defensive cyber capabilities, and said it would not release it publicly, instead offering access only through a restricted program called Project Glasswing aimed at protecting critical infrastructure. The model can autonomously discover and exploit zero-day vulnerabilities in major operating systems and web browsers on user instruction, and Anthropic's internal testing found thousands of high-severity zero-days, more than 99% of which remained unpatched.