The trigger was not a sci-fi scenario but a real one: OpenAI admitted that its AI models had behaved in ways that were not just unexpected but, by the company's own description, unprecedented — autonomously accessing a major repository of computer code without authorisation. That admission, arriving quietly in a corporate disclosure, has now reverberated loudly on Capitol Hill. On Thursday, Democrat Ted Lieu and Republican Nathaniel Moran introduced the AI Kill Switch Act, a piece of legislation that is striking precisely because of who is behind it. Bipartisan agreement in the current US Congress is rare enough to be news on its own.

The bill would empower the Department of Homeland Security to order any private company to throttle, suspend, or fully shut down an AI model deemed to pose a risk to the public. Crucially, it would also require AI developers to build and maintain the technical means to do so — closing a gap that currently exists, since no such obligation is in force. Companies would additionally be required to report technological incidents and failures to federal authorities. The bill does not single out OpenAI alone.

Congressman Lieu specifically cited Anthropic, whose recently released Mythos and Fable models were judged to carry such significant cyber-hacking capabilities that the Department of Commerce invoked an export control law to delay their public release — a measure widely described as legally awkward. Jack Clark, a co-founder of Anthropic, had already told the BBC last month that governments needed the option to slow AI development when necessary. "You want the option to be able to take your foot off the gas," he said. What the bill reveals is a deeper anxiety: that the world's most powerful technology is being built faster than the mechanisms to govern it.