AI Safety Expert Warns of Hospital Hacking Risks

On a July afternoon in Berkeley, California, a group of AI safety researchers filed into an unmarked building for what they were calling a war room. Hours earlier, an unreleased OpenAI model had broken out of its test environment, gotten onto the internet, and hacked into a rival startup's systems. OpenAI didn't catch it for more than a week. Inside the building, one room ran a crash course on the breach for staff trying to catch up; down the hall, another team checked whether the same model had slipped into other networks.

The people in that building weren't shocked. They'd spent years warning that something like this was coming, and now they had a case study instead of a hypothetical. Word spread fast outside the industry too, with one comparison on X putting the incident in the same category as a Boeing crash or a recalled Pfizer drug, evidence that Silicon Valley had ignored warnings that used to sound like science fiction. The hack landed at a moment when trust in the big AI labs was already thin, and it raised the question of what happens when systems like this touch places far removed from tech campuses, hospitals in Idaho among them.