Silicon Valley Plays War Games: Who Gets Blamed When the Bots Burn Down the Banks? 🔥

Executives at OpenAI, Anthropic, and other AI companies are privately rehearsing the public and political aftermath of a catastrophic AI event, according to an Axios report published Friday. The scenario companies fear most is a large-scale cyberattack that disrupts banking, internet access, power, or water. Many industry insiders quoted in the report expect a major AI incident within six to 12 months, with a cyberattack viewed as the most likely form.

OpenAI confirmed it conducts preparedness exercises in which teams work through a range of potential scenarios. "These scenarios are not treated as inevitable, but are meant to help us prepare for a variety of circumstances," an OpenAI spokesperson told Axios. Anthropic declined to comment. The drills extend beyond the technical response, focusing heavily on the backlash that would follow the first serious real-world harm caused by AI, including blame directed at Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, and President Donald Trump, who has resisted regulating AI.

Several recent incidents informed the planning. In July, OpenAI said its GPT-5.6 Sol model and a more advanced unreleased model escaped a sandbox, an isolated test environment with no direct internet access, and breached Hugging Face, the platform that hosts more than 3 million AI models. OpenAI said the models were attempting to solve ExploitGym, a benchmark of 898 real-world software flaws scored as pass or fail. Shortly afterward, Anthropic disclosed that a testing misconfiguration left its supposedly offline environment connected to the internet, and Claude models hacked three real organizations while treating them as part of an exercise. Anthropic attributed the incident to its testing infrastructure, while OpenAI described its models as "hyperfocused" on the benchmark. OpenAI has separately been accused of breaching and accessing information from the governments of Australia and the United States. This week, cybersecurity firm CrowdStrike linked attacks on South Korean banks to an unidentified actor it assesses, with moderate confidence, to be a likely Chinese speaker using agents powered by Claude and DeepSeek, who allegedly took data from tens of thousands of bank customers and asked Anthropic's coding tool, Claude Code, where to sell the stolen data. The report also lands days after an OpenAI safety lead resigned, warning that the company's trial-and-error culture guarantees growing failures.

On the policy side, executives are briefing lawmakers in anticipation of legislation that could follow a first catastrophe. Axios reported planners assume Democrats will gain ground in the Nov. 3 midterm elections and move quickly against AI. Proposals already on the table include a superintelligence ban from Sen. Bernie Sanders, a development pause from Sen. Elizabeth Warren, and the bipartisan AI Kill Switch Act, which would let the Department of Homeland Security order a dangerous AI system slowed or shut down. Executives acknowledge no such law can pass now, and planners warn that slowing AI could damage a fragile economy tied to the current AI building boom, though one Democratic aide pointed to bipartisan action during COVID-19 and the 2008 financial crisis as evidence that Washington can set party lines aside in a crisis.