The White House Is Making Up Its Rules for AI in Real Time

## Background of the Dispute It has been almost a week since the Trump administration issued an export‑control directive that forced Anthropic, one of the world’s leading AI laboratories, to take its most advanced models—Claude Mythos and Fable 5—offline. After several days of back‑and‑forth negotiations, the two parties remain at an impasse over how, or even if, the models can be reinstated.

## White House Position According to a former White House technology official who asked to remain anonymous, the administration views Anthropic’s actions as reckless and believes the company cannot be trusted to safely deploy frontier AI. The official criticized the “extreme anti‑regulatory posture” of the current administration, arguing that the lack of prepared policies forces a “slap‑dash” approach that puts the entire AI sector in a quandary.

## Anthropic’s Response A source close to Anthropic maintains that the firm did not breach any concrete procedures or rules laid out by the Trump administration. The company asserts that it has complied with all known export‑control requirements and that the White House has not provided a clear explanation of the alleged violation. The only public statement from the administration comes from White House technology adviser David Sacks, who posted a brief outline of the situation on X.

## Concerns Over International Partnerships U.S. officials grew uneasy earlier this month when they learned that Anthropic had shared Claude Mythos with SK Telecom, a South Korean telecom giant they allege has ties to China. Anthropic says the collaboration had been ongoing for years without prior national‑security concerns, and that it coordinated with the government during the rollout. When the White House raised the issue, Anthropic promptly revoked SK Telecom’s access.

## Technical Challenges with Jailbreaks Amazon CEO Andy Jassy raised additional worries to Treasury Secretary Scott Bessent about potential circumventions of the guardrails on Claude Fable 5, a safeguarded version of Mythos. Both Anthropic and independent cybersecurity researchers note that “jailbreaking” large language models is an inherent, complex problem. Because AI models are probabilistic rather than deterministic, developers cannot guarantee exact outputs for any given prompt, making a complete fix unlikely in the short term.

## Implications for the AI Industry The episode illustrates what many observers now describe as the “Wild West” era of American AI regulation. While few statutes currently govern frontier AI, the Trump administration’s ad‑hoc actions create de‑facto licensing expectations. The administration’s recent executive order established a “voluntary” submission system for AI labs to provide early‑stage models for government testing, but critics argue that, in practice, it functions as a mandatory licensing regime.

## Future Outlook AI leaders at OpenAI, Google, Meta, and other labs are watching Anthropic’s situation closely. Cohere CEO Aidan Gomez says the industry is moving toward “advance notice, advance access” with the government to avoid surprises. The prevailing sentiment is that proactive engagement with the White House may become a prerequisite for releasing cutting‑edge models without triggering further export‑control actions.

This article is part of Maxwell Zeff’s Model Behavior newsletter. Previous editions can be read on the newsletter’s archive.