Claude Mythos Preview: The AI That Can Hack Everything (And Why You Can't Use It)

Anthropic just published a 244-page system card for a model they have
zero intention of releasing publicly.

The model is called Claude Mythos Preview. And the reason you can't
use it isn't pricing or performance — it's because they believe it's
too dangerous.

Here's what it actually did in testing:

Found a 27-year-old bug in OpenBSD — autonomously, in hours.
OpenBSD is one of the most security-hardened OS projects in existence.
Security researchers reviewed its code for nearly 3 decades. Mythos
found a remote crash vulnerability without human steering.

Found a 16-year-old bug in FFmpeg.
Automated tools had hit this codebase 5 million times. Nobody caught it.

Chained multiple Linux kernel zero-days to escalate from a normal
user to full machine control.

Anthropic's own Red Team researcher said:
"I've found more bugs in the last couple of weeks than I found in
the rest of my life combined."

What is Project Glasswing?

Instead of releasing Mythos publicly, Anthropic launched Project
Glasswing — giving access only to 12 organizations (AWS, Apple,
Google, Microsoft, CrowdStrike etc.) for defensive security work.

They've committed $100M in usage credits for this initiative.

Should developers be worried about their jobs?

That's the real question. Mythos isn't a narrow security tool —
it's a general-purpose model that happens to be extraordinary at
finding vulnerabilities.

I did a full breakdown covering: