Inside the White House's Regulatory Game with Anthropic

U.S. political news site Politico published an in-depth investigative report, interviewing over a dozen parties, revealing for the first time the complete account of the 85-day regulatory battle between the Trump administration and Anthropic regarding the safety regulation of cutting-edge AI models from April to July this year. This confrontation essentially laid the groundwork for the current regulatory framework for cutting-edge AI models in the U.S.

The trigger was Anthropic's release of the new model Mythos to approved users on April 7—this model can quickly identify security vulnerabilities in digital infrastructure, serving as a powerful tool for defense and an automated weapon for attack; after Microsoft gained early access at the end of March, it privately alerted the government, prompting Treasury Secretary Bessent to convene the CEOs of the five major banks overnight, leading the White House to activate its crisis response mode.

White House Chief of Staff Wiles appointed National Cyber Director Cairncross, who had a background in election law and was nearly unfamiliar with cybersecurity, to lead the development of regulatory policy. His team issued a 31-question survey to companies, and a draft released in early May required a 60-day mandatory assessment by the Commerce Department's CAISI agency before the release of cutting-edge models, which faced collective opposition from Google, Anthropic, and OpenAI. Internal documents from OpenAI described this window as "operationally unrealistic."

On May 21, the day the executive order was originally scheduled to be signed, former AI czar Sacks called Trump, persuading him to temporarily cancel the signing on the grounds that "slowing down will help China"; the executive order was ultimately signed quietly on June 2, with the review window compressed from 90 days to 30 days and made voluntary, with no ceremony or cameras involved.

On June 9, Anthropic released the public version of the model Fable. Two days later, Amazon researchers discovered a jailbreak vulnerability, and the White House demanded an immediate takedown, which was refused by Amodei. The two sides engaged in a remote confrontation—Cairncross stated, "We have gotten the answers we need," while Treasury Secretary Bessent bluntly said, "This is a bad decision"; that afternoon at 5:21 PM, the Commerce Department issued an export control directive prohibiting foreign nationals from accessing it, effectively equivalent to a global takedown due to Anthropic's own overseas engineers.

Anthropic co-founder Tom Brown was urgently appointed to take over negotiations, and within 24 hours, he and his technical team flew to Washington to demonstrate safety improvements with real-time data and trained an improved safety classifier. Nineteen days later, Fable was restored online; after this incident, OpenAI adopted a more cautious approach when launching new models, proactively accepting phased release arrangements, while telecommunications infrastructure company Telnyx shifted to China's Zhizhu GLM model to avoid similar risks.

In September, Amodei publicly called for "slowing down the cutting-edge pace," receiving public support from Altman and Musk, while Trump referred to AI risks as a "scam"; the executive order's safety review details, which were to be published by the end of August, have yet to be made public. A policy-making official summarized: "There is a hammer outside, and companies know what it is."

Source: Public information