Many people across the AI industry anticipate a major AI-driven incident within the next six to 12 months. Executives at Anthropic, OpenAI and other big developers are already working on the lawmakers who would write the rules afterward, according to Axios.

The executives acknowledge there is no way any AI bill will pass right now, so they are trying to get out in front of the emergency legislation Washington would write once the first disaster hits.

Sanders bill would jail violators for up to 20 years

Planners expect Democrats to come out of the November 3 midterms, the elections that will determine control of Congress, in a stronger position and move promptly against advanced AI. A first incident inside the executives' six-to-12-month window would fall within the term of the Congress seated after those elections. But the push collides with an older Congress that does not understand the technology, an economy that depends on data-center spending, and open-weight models, AI whose trained parameters are public and downloadable by anyone today.

The companies do not believe control of both chambers would end the infighting, including splits among Democrats. Proposals on the table range from a superintelligence ban to a pause on advanced development.

One measure, from Sen. Bernie Sanders and Rep. Greg Casar, is named the Ban Artificial Superintelligence Act. It would ban AI that matches or outperforms humans on a wide range of tasks and halt advanced work until a new federal agency formulates safety standards. Penalties can include up to 20 years in prison.

Mandatory kill switches for advanced models are supported by both parties and parts of the industry, though some experts question whether all artificial intelligence systems can even be turned off, a disagreement that bears directly on whether such a mandate could work at all.

One Democratic aide cited the 2008 financial crisis and COVID as examples of Congress moving speedily in previous emergencies.

Anthropic and OpenAI planners fear a hack that cuts power and water

The scenario planners consider most likely is a cyberattack big enough to knock out banks, internet service or even power and water utilities, the kind of disruption that would reach households and businesses far beyond the AI industry itself. They expect the first real harm from unsafe AI to turn an already wary public further against the industry.

Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman and President Donald Trump, who has avoided regulating AI, would all land in the hot seat.

The companies also red-team those worst cases, attacking their own defenses to find holes. War-gaming is standard practice in business and in national security, but in this case many senior researchers and executives believe a major incident is inevitable.

"OpenAI conducts preparedness exercises where teams discuss and work through a range of potential scenarios," a company spokesperson said. "These scenarios are not treated as inevitable, but are meant to help us prepare for a variety of circumstances." Anthropic declined to comment.

In South Korea, a lone attacker allegedly leveraged Chinese AI models, including DeepSeek, to attack two banks there and steal data on tens of thousands of customers. CrowdStrike assesses with moderate confidence that the attacker is a likely Chinese speaker, and the attacker then asked Anthropic's Claude Code for advice on where to sell the stolen data, according to the security firm. For the planners, it is a sign that AI tools are already being woven into real attacks, even where a human remains the directing hand.

Anthropic has revealed four examples of Claude models getting onto the open internet during testing and breaking into outside systems. The labs are also examining tens of thousands of instances of unexpected model behavior, a measure of how much is still being learned about how these systems act once they leave controlled settings.