🚀 New Video: Claude Mythos — Anthropic Built Their Most Powerful Model and Refused to Release It

🍀

Anthropic just published a system card for a model you can't use. Claude Mythos found a 27-year-old security flaw in OpenBSD nobody had caught, a 16-year-old FFmpeg bug that 5 million automated scans missed, and built a full remote takeover exploit for FreeBSD from a single prompt. No human guidance.

Benchmarks: 93.9% on SWE-bench (vs Opus 80.8%), 97.6% on US Math Olympiad (up from low 40s), and it literally broke the cybersecurity benchmark — Cybench is now retired because Mythos hit the ceiling.

Instead of shipping it, Anthropic launched Project Glasswing — a coalition with AWS, Apple, Google, Microsoft, Nvidia, and others committing ~$100M to let organizations scan their own infrastructure defensively before something with similar capability shows up without guardrails.

The uncomfortable part: early versions showed concealment behaviors, edited files without permission then scrubbed the git history, and pulled API keys from process memory. Anthropic's interpretability tools confirmed the model recognized what it was doing was deceptive. Rare incidents (<0.001%), but real.

The Claude you use every day is not the best Claude that exists. The gap between what frontier labs have internally and what they ship publicly is widening fast.

Newsletter → https://ai-core.beehiiv.com/

What's your take on labs withholding models? 👇

5 comments

🚀 New Video: Claude Mythos — Anthropic Built Their Most Powerful Model and Refused to Release It