Claude Mythos Preview: a model that finds zero-days at industrial speed, and Project Glasswing
In one sentence Anthropic announces Claude Mythos Preview: a model with extraordinary cyber capabilities (thousands of zero-days identified across OSes and browsers, 181 working Firefox exploits). Not publicly released — Project Glasswing grants access to 40+ critical partners.
On April 7, 2026 Anthropic announces a model that radically changes the conversation about cybersecurity. It's called Claude Mythos Preview. They are not releasing it to the public.
Why? Because Mythos is extraordinarily good at finding vulnerabilities in existing software. In internal tests Anthropic identified thousands of zero-day vulnerabilities (never-before-disclosed) across major operating systems and web browsers. In a focused Firefox test, Opus 4.6 found working exploits 2 times in several hundred attempts; Mythos found them 181 times. It doesn't merely describe vulnerabilities — it writes working exploits, chains multiple bugs into real attacks, reconstructs source from closed binaries.
Anthropic launches "Project Glasswing": instead of releasing Mythos to everyone, they give it to ~40 critical partners (Apple, Microsoft, Google, Amazon, NVIDIA, Cisco, CrowdStrike) to harden their infrastructure before similar capabilities become available to attackers. Anthropic estimate: 6-18 months until other labs reach equivalent capability.
On the same day the model is announced, a small unauthorized group gets access to Mythos via a third-party vendor. The AI safety debate moves from theoretical to operational.
Companies
Anthropic
Tools
Claude Mythos Preview
Tags
Sources