CEO of Anthropic:
“We haven’t trained it specifically to be good at cyber, we trained it to be good at code. But as a side effect of being good at code it’s good at cyber too.”
Now the most powerful cybersecurity AI model ever built just got breached. Anthropic says there’s no evidence it affected their systems, but that’s not really the point.
The interesting part is that a system meant to secure everything ended up having a basic access control issue itself.
NEW: A small group of "unauthorized users" have reportedly breached Anthropic's tightly restricted Claude Mythos.