Anthropic built their most powerful AI model ever, Mythos, and decided not to release it to the public.

During internal testing, it found previously unknown security vulnerabilities, the kind that give attackers a way in before anyone knows to look, across every major operating system and browser. It cracked open a 27-year-old flaw in OpenBSD, one of the most security-hardened operating systems on the planet. It found a hole in FFmpeg, a video processing tool embedded in billions of devices, that had survived five million automated security scans without being detected.

Anthropic engineers with zero security backgrounds pointed it at software overnight and woke up to working exploits.

So they gave access to Amazon, Apple, Microsoft, CrowdStrike and a handful of others. No public release.

I’ve been doing this work long enough to know that the technology is rarely the real story. The real story is always about people. Any geek could always learn computer forensics, but having an investigator in computer forensics is different. And learning Microsoft Word has yet to make me into Hemingway.

AI this capable is only as good as the people using it. Access isn’t the differentiator anymore. Proficiency is.

HUMINT first!