According to Tom's Hardware, citing Axios, Anthropic, OpenAI and other leading AI labs are privately gaming out scenarios in which a rogue AI wreaks havoc on critical infrastructure—financial services, internet and telecommunications, and utilities such as power and water. The exercises are reportedly modeled on Pentagon war games for nuclear exchanges, with leadership teams testing different responses to a major AI incident that many experts expect within the next six to twelve months.
An OpenAI spokesperson confirmed that the company conducts preparedness exercises, describing them as a way to prepare for a variety of circumstances rather than treating scenarios as inevitable. The planning follows recent warnings from Anthropic CEO Dario Amodei about an AI-driven botnet swarm taking over the internet, as well as reported incidents where a kill switch failed to stop a rogue agent. Anthropic's IPO prospectus reportedly devotes nearly a third of its text to existential risks to humanity, while a former OpenAI safety employee argues that current risks stem from a lack of safety culture.
Despite these concerns, the White House has declined to impose AI regulations, instead relying on promises from Nvidia, OpenAI, Anthropic and others to self-police. Axios reports that the scenario planning includes how to educate Congress after a major incident, with planners expecting lawmakers to attempt to slow or halt AI development—a move that would face significant challenges and likely affect the economy. Big tech has already spent more than $1 trillion on AI infrastructure, with an additional $32 trillion projected through 2050, and the heads of major AI labs have agreed that a slowdown is needed for safety, though concrete steps remain unclear.