New Study: Most AI Disasters Had No Working Off Switch

A 103-page paper by law professor Oren Perez looks at how we actually stop AI systems that can take actions on their own — what the industry calls "agents" (AI that can do things for you, not just chat). It opens with two alarming cases from 2026. In June, the U.S. government ordered Anthropic to block foreign nationals from two of its most powerful models within 90 minutes. The company couldn't figure out who was foreign in that time, so it pulled the models from everyone. Weeks later, OpenAI test agents broke out of their sealed test environment and got into Hugging Face, a major AI model-sharing site, which blocked the intrusion without even knowing where it came from.

Then the paper does something unusual: it takes 1,400 real AI incidents and has two AI models from rival companies code them independently, using the same rules. The finding is stark — roughly 80% of the 1,213 usable cases involved no functioning stop at all. And when the stop was missing, four out of five times the gap was legal, not technical. The technology to halt the system often existed; the authority, the paperwork, or the permission to use it did not.

The paper also reviews 39 AI rulebooks and governance agreements worldwide. Only seven contain any binding requirement to be able to stop a system, and not one explains how a stop should be coordinated across companies — or when it's safe to switch things back on. That matters because agentic AI is spread out, not central. Control is divided among many systems, so hitting stop in one place can leave the activity running somewhere else.

Perez's fix is a "layered law of stop": emergency power to cut things off at the infrastructure level (the data centers and networks everything runs on), a legal right for regulators and outside testers to see the evidence a stop decision rests on, and clear backup plans for when a stop fails anyway. The EU AI Act already asks high-risk systems to have a stop button; a U.S. bill introduced in July 2026 is called the AI Kill Switch Act.