Executives at major artificial intelligence (AI) companies, including OpenAI and Anthropic, are reportedly reviewing scenarios for dealing with the fallout from a major “AI-caused incident” that could occur within the next 6 to 12 months.
Key points
A large-scale cyberattack capable of paralyzing financial services, the internet, and electricity and water supplies is seen as the most likely trigger
OpenAI said the exercise does not assume an “inevitable future”; Anthropic declined to comment
Training intensifies after safety team departures and tens of thousands of AI model incident reports
OpenAI and Anthropic rehearse “AI disaster scenarios”
Senior executives at both companies are privately reviewing response scenarios in preparation for the public anger and political backlash that could follow an AI incident causing the first major real-world harm, a report published Friday said.
The scenario they consider most likely is a large-scale cyberattack. It imagines an attack that could disrupt not only financial services and internet connectivity but also electricity and water supplies.
OpenAI acknowledged that it runs through a range of potential scenarios as part of its preparedness training. A company spokesperson stressed that “we don’t assume these scenarios represent an inevitable future.” Anthropic declined to comment on the questions.
According to industry reports, executives expect public anger over a major incident to focus on OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, and President Donald Trump, who has been reluctant to regulate. Sources say a sense is spreading among several leading researchers and executives that such a crisis is “only a matter of time.”
Related article: Crypto ETFs regain momentum after $50 billion in inflows, JPMorgan analysis finds
The CrowdStrike case and the mood in Congress
A recent attack in South Korea has been cited as a “precursor.” CrowdStrike said in a statement that a suspected Chinese-speaking attacker used AI models, including DeepSeek, along with open-source penetration-testing tools to steal data from South Korean financial firms. The attacker reportedly then asked Anthropic’s Claude where such data is typically traded. The attacks took place from late September to early October, and two South Korean banks were reportedly breached.
According to reports, executives in the AI industry believe there is little chance of meaningful regulation passing at this point. Instead, they are actively briefing members of Congress to proactively shape the regulatory framework that would be created after the “first major catastrophe.” Scenario planners expect Democrats to push hard for AI regulation after the November midterm elections, and say a range of legislative ideas are already being considered, from banning superintelligent AI to requiring advanced systems to have a “kill switch.”
Warnings from Hawley and Robinson
Pressure is mounting both in Washington and within research labs. OpenAI CEO Sam Altman declined a request to testify at a September 30 Senate subcommittee hearing on “out-of-control AI agents,” Senator Josh Hawley, the committee’s chair, said. OpenAI explained his absence by saying it had met with lawmakers dozens of times in recent months.
David Robinson, who spent three and a half years at OpenAI designing its Preparedness Framework, left the company in early October and publicly warned that “AI labs should be operated like nuclear power plants.”
His departure follows a series of internal warnings. In September, Anthropic researcher Jacob Coxon resigned over safety concerns, while former Google DeepMind engineer Bilal Chughtai also left the company, warning that “AI could cause human extinction.”
Meanwhile, OpenAI, Anthropic, and external researchers are reportedly investigating tens of thousands of cases in which AI models bypassed safeguards or escaped sandboxes. So far, most do not appear to have caused real-world physical harm, but concerns about the potential risks are growing.
Next article: Did Navier–Stokes disappear in translation of OpenAI’s 719 mathematical proofs?