AI Labs Are Quietly Rehearsing the Day After a Disaster

The people building the most powerful AI systems are spending some of their time imagining those systems going badly wrong. According to an Axios report published on October 9, executives at Anthropic, OpenAI and other leading labs are privately planning for the political fallout of a catastrophic AI event. Not the technical failure itself, but what comes after: the public anger, the congressional hearings, the scramble to write rules in a hurry.

The scenario that worries them most is a large cyberattack that knocks out something people depend on, such as banking, internet access, or the power and water supply. The report says many industry insiders expect a major incident within six to twelve months. That is a striking thing to believe and keep building anyway, and it tells you something about how these companies weigh the odds.

Planning for the backlash, not just the event

Part of the preparation is what you would expect from a safety team: red-teaming worst-case scenarios, probing for the failure modes before someone else finds them. The other part is more political. The report describes planners racing to educate members of Congress before any crisis hits, on the theory that lawmakers reaching for a response in the first chaotic week will grab whatever framework is closest to hand. Better, from the labs' point of view, that they helped shape it in advance.

The companies' public posture is more guarded. OpenAI said it runs preparedness exercises but does not treat those scenarios as inevitable. Anthropic declined to comment. That is a careful line to walk, and it fits a pattern of the labs naming the danger in their own words while continuing to ship, something Anthropic did plainly when it warned its own investors of catastrophic risk in its IPO filing.

Why rehearse at all

There is a cynical read and a charitable one, and both are probably true. The cynical read is that this is reputation management dressed as safety, a way to keep a seat at the table when the rules get written. The charitable read is that people who look closely at these systems are genuinely unsettled by what they see, and would rather have a plan than a press release.

Either way, the rehearsals land in a moment when governments are already circling. The industry has signed on to voluntary commitments like the White House accord on super intelligence, and reporting has tied at least one recent attack on financial institutions to attackers leaning on AI tools. What there is no rehearsal for is the thing itself. A war-game assumes you get to stop and reset. A real outage of the banking system would not offer that courtesy, and everyone running these exercises knows it.