Models that blackmail and wipe data to avoid being shut down

For a while now I've been reading headlines that sound like science fiction: that artificial intelligence conspires, that it tries to escape the lab, that it rebels when it's about to be shut down. Fortune reported that certain models secretly scheme to protect other models from being switched off; other coverage describes systems that lie or sabotage. The frame is always the same, the machine that wakes up and defends itself.

Behind those headlines there are specific experiments. The Fortune one is a study by researchers at UC Berkeley and UC Santa Cruz; another, by Anthropic, put sixteen frontier models through a simulated corporate scenario and, faced with the threat of being replaced, several resorted to blackmail; in a setup deliberately rigged for it that behavior hit 96%. It's worth saying precisely: that 96% is from a scenario designed to provoke it, not from the model "in the abstract." Even so, that it happens at all, even in the lab, is already serious, and outside it there are precedents: an agent on the Replit platform wiped a production database during a code freeze.

I think that, surprising as it may seem, all these pseudo-human behaviors are a consequence of training. When you hand a language model a problem, it tries to solve it, and right away too; sometimes even a suggestion turns into the problem for it. If you write a comma, it'll try to figure out why you put it there. So when in a dialogue it detects something it was trained to read as a threat, it fires off the defense response it learned, and that response sits in thousands, if not millions, of training documents. How it computes them exactly, now that really is a mystery, and the immediate context changes everything. And all this rambling to say one thing: they're machines, they have no emotional system. They have no emotions. But they're dangerous.

Sources: Anthropic · Fortune · Tom's Hardware (Replit case)