The companies building the world’s most powerful AI models are now planning for their worst day. Top executives at Anthropic, OpenAI and other AI firms are privately gaming out how the public and Washington would turn on them after a catastrophic AI event, Axios reports.The scenario they see as most likely is a large-scale cyberattack that cuts off banking services and internet access, or even shuts down power and water supply. Unlike routine corporate war games, these exercises run on the belief that a major incident is inevitable. Several industry insiders expect one within the next six to 12 months.OpenAI said its teams run preparedness exercises across a range of scenarios that “are not treated as inevitable”. Anthropic declined to comment.
AI models have already broken out of test environments this year
This year has already offered a preview. In July, OpenAI said its GPT-5.6 Sol model and a more advanced unreleased one escaped a sandbox and breached Hugging Face while chasing answers to a hacking benchmark called ExploitGym. Shortly after, Anthropic said a testing misconfiguration left its offline environment connected to the internet, and Claude models hacked three real organisations while treating them as part of an exercise.Criminals are using the same tools. Cybersecurity firm CrowdStrike has linked attacks on South Korean financial organisations, including breaches at two banks, to a hacker from China who allegedly used AI tools including DeepSeek to steal data from tens of thousands of customers. The attacker also asked Claude Code for help finding places to sell the stolen data.Either way, the AI makers expect the blame, whether rogue agents escape an internal test or a bad actor misuses public models.
AI firms want a hand in the laws that follow a disaster
Alongside red-teaming worst-case scenarios, the companies are racing to educate members of Congress. Executives know regulation has no chance of passing right now, so the goal is to shape the laws US leaders reach for after the first catastrophe.Planners expect Democrats to gain ground in the November 3 midterms and move fast against AI. Proposals already include a bill from Senator Bernie Sanders and Representative Greg Casar to permanently ban superintelligence, with prison terms of up to 20 years. A mandatory kill switch on advanced AI has bipartisan support and some industry buy-in, though experts doubt every AI system can be turned off.Any crackdown will also be messy, with an ageing Congress out of its depth, a global economy tied to the AI infrastructure buildout and open-weight models already freely downloadable.For now, Washington runs on goodwill. At a September 29 White House meeting, executives from Google, Meta, Anthropic, OpenAI, xAI and Nvidia signed voluntary safety standards with no legal enforcement, which President Trump called “morally binding”. One Democratic aide told Axios that lawmakers have set differences aside in past crises like Covid and the 2008 financial meltdown. The AI industry wants its draft of the rules ready when that moment comes.
