Anthropic's Mythos Model: Near-Superhuman at Security and Math, Coming Soon to Public

Mythos (unveiled April 7, currently restricted to ~50 defensive-security partners via Project Glasswing) scores 97.6% on the 2026 USA Mathematical Olympiad vs 42.3% for Opus 4.6, and successfully exploits Firefox 147 vulnerabilities in 181 of several hundred attempts vs 2 for Opus 4.6. The UK AI Security Institute reports 73% success on expert-level hacking tasks that no model could complete before April 2025. Anthropic plans a general release once stronger safeguards are in place — likely paired with an upcoming Opus model. Developers building security tooling should watch this closely.

---

Sources: The Register · Anthropic Glasswing