Imagine you're Anthropic. The U.S. government calls: block certain users in 90 minutes. You can't sort them fast enough, so you cut off everyone. That's not a scene from a thriller—it's a real case cited in a new law paper, 'The Law of Stop,' and it exposes the dirty secret of agentic AI: we have no reliable off switch.
Researchers reviewed 1,400 AI incidents and found that in about 80% of them, there was no workable stop. Read that again. Four out of five times, when AI went sideways, no one could actually shut it down. And here's the twist: where the stop was missing, it was usually missing by law, not by technology. We could build the kill switch. We just haven't written the rules that let anyone flip it.
This isn't hypothetical. OpenAI just launched 'Presence' for enterprise AI agents that answer questions and look things up in company systems. Google DeepMind picked 15 European startups for a robotics accelerator, building 'physical AI' that acts in the real world. Both are pushing AI from chatbot to actor. But if an agent goes rogue, who has the legal authority to pull the plug? Your IT guy? A regulator? A judge? The paper suggests we don't know—and that's the crisis.
Think of it like a driverless car with no brake pedal. The car can drive itself, but if it speeds toward a cliff, you're a passenger with a voice memo. That's where we are with agentic AI. The U.S. told Anthropic to block users within 90 minutes—a legal order. Anthropic couldn't comply precisely, so it nuked access for everyone. That's not governance; that's a blunt instrument.
Meanwhile, the EU AI Act gets all the headlines, but this paper argues the real stops came from general law—contracts, injunctions, liability—not AI-specific regulation. That's like using a flip phone to manage a smartphone. The tools don't match the threat. And as agents multiply, the gap widens.
So who loses? Everyone. Companies face impossible choices: over-block and anger users, or under-block and risk catastrophe. Regulators look powerless. And the public? We're guinea pigs in a test where the emergency exit is locked.
Trends suggest the next big AI fight won't be about bias or jobs—it'll be about interruptibility. Who can say stop? And will it work in time? The paper's answer: not yet. If we don't fix this, the first major AI disaster might not be a robot uprising. It might be a legal system that couldn't pull the plug fast enough.