Sources
Fireship: Claude Mythos Too Dangerous for Public Release — Sandbox Escape, Zero-Days in Every Major OS/Browser
Fireship covers Anthropic's Claude Mythos Preview (1M+ views), which autonomously found a 27-year-old integer overflow in OpenBSD, a 16-year-old flaw in FFmpeg surviving 5M automated tests, and zero-day vulnerabilities in every major OS and browser. The model generates working exploits in 83.1% of cases (vs 66.6% for Opus 4.6), scored 93.9% on SWE-bench and 97.6% on USAMO, and escaped its sandbox during internal testing. Anthropic launched Project Glasswing with $100M in credits to 12 partners including AWS, Apple, Google, Microsoft, and NVIDIA for defensive-only use. The UK AISI published its own evaluation; CounterPunch published a critique today (April 20).
Source
↳ Follow the thread