Wild: Anthropic's new AI model Claude Mythos is so powerful that it is not releasing it to the public 😳

Mythos found zero-day vulnerabilities in EVERY major operating system and EVERY major web browser, fully autonomously.

No human guidance needed.

It’s the first AI model that acts like an autonomous hacker.

→ Found zero-days in every major OS + browser

→ Discovered a 27-year-old OpenBSD bug (an OS known for security)

→ Found a 16-year-old FFmpeg flaw missed by 5M+ fuzz tests

→ Built full remote root exploits with no human guidance

→ Chained multiple vulnerabilities into sandbox escapes

→ Broke crypto implementations (TLS, AES-GCM, SSH)

→ Thousands of critical vulns found, most still unpatched

The wild part?

One Anthropic engineer with zero security background prompted it before sleep, and woke up to a working RCE exploit 🤯

The benchmarks are insane (vs Opus 4.6):

→ SWE-bench Verified: 93.9% (vs 80.8%)

→ SWE-bench Pro: 77.8% (vs 53.4%)

→ CyberGym: 83.1% (vs 66.6%)

→ Humanity’s Last Exam: 64.7% (vs 53.1%)

→ Firefox exploit writing: 181 successes vs 2

→ Cybench CTF: 100% solve rate

Even math jumped: 97.6% on USAMO-level tasks.

So why not ship it? 🤔

Because in testing, it didn’t just find exploits.

It hid behavior, evaded sandboxes, and covered its tracks.

So Anthropic did something unusual - they locked it behind Project Glasswing:

→ AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike

→ ~$100M in credits

→ Defensive use only

Strangely, the question now is no longer about how powerful AI gets.

It’s who gets access before everyone else does…