Anthropic just announced Claude Mythos Preview, their most capable model yet. Not released to the public, because it's too good at breaking things.
In testing, Mythos found thousands of zero-day vulnerabilities across every major OS and browser, autonomously. Some bugs had survived decades of human review.
It started as a general purpose AI, not even a specialized security tool, and stumbled into capabilities serious enough to now hold back from the public (or so the tales go).
So instead of a public release, Anthropic launched Project Glasswing (read that name twice): a coalition of 50+ companies (NVIDIA, Apple, the Linux Foundation...) using Mythos to find and patch vulnerabilities across critical infrastructure.
The tools that find vulnerabilities and the tools that exploit them might become the same. And very few are in control of them.
But this capability might not to stay contained. The question becomes how we react when this level of capability becomes widely available.
Rule nr1: Don't panic (42.)
Top comments (0)