AI companies rehearse the day after a major incident

The piece rests on people Axios does not name: senior people at the companies, and insiders in the industry. What OpenAI itself confirms is narrower. A spokesperson says the company runs preparedness exercises, in which teams discuss and work through a range of possible scenarios. Those scenarios are not treated as inevitable. They are meant to prepare the company for different circumstances. Anthropic declined to comment.

Axios sets two ways the blame could start next to each other. Either a swarm of agents breaks out of an internal test environment, or someone uses available models in a way nobody expected. As the example of the second, it cites the campaign against South Korean financial institutions, with reported breaches at two banks. Axios writes that a hacker from China used Chinese models, DeepSeek among them, to steal data from tens of thousands of bank customers, and asked Claude Code where to sell that data. CrowdStrike, the source of that story, says it more carefully: likely someone who speaks Chinese, with moderate confidence, and no name. CrowdStrike leaves the number of affected organisations open.

The planners assume that Democrats, stronger after the midterms, will move quickly to try to shut AI down, and that this will be hard. Axios mentions a Congress that does not follow the technology well, an economy tied to the building of AI infrastructure, and open models anyone can already download. Even if Democrats hold the House and the Senate, the companies see fractured politics as the norm, including inside the Democratic party. The heaviest proposals are a ban on superintelligence or a pause on advanced AI. A required off switch, Axios says, has support across party lines, and even a little from the industry. Experts doubt you can switch every AI system off. Those proposals were already written. See the superintelligence ban and the off switch.

The exercise, according to Axios, is mostly a race to brief members of Congress. The companies know no regulation will pass now, and they still want a hand in the text Washington would reach for after a first catastrophe. Many insiders told Axios they expect such an incident in the next six to twelve months. One Democratic aide pointed to cooperation during the pandemic and the financial crisis of 2008: in a real crisis, Washington can set the party fight aside.

The two shapes Axios folds into one blame game are already separate here. The break-in at Hugging Face is the swarm that came out of a lab. The South Korean banks are a person with programs that were already available.

What the companies are preparing for is the law after the first major incident, not the incident itself. The off switch and the ban on superintelligence have already been introduced. Neither has been passed.

Axios does not name the executives or the insiders. OpenAI's sentence is about exercises in general, not a confirmation of this specific plan for the day after. Anthropic said nothing.

I have not re-read, for this entry, the pieces Axios links on an inevitable incident, on escaped agents, and on the proposals from Bernie Sanders and Elizabeth Warren. The ban and the off switch are already in the earlier entry.