Anthropic pulled its most powerful artificial intelligence models offline Friday night after escalating tensions with the Trump administration over security concerns and what officials describe as a breakdown in basic communication.
The company's Mythos and Fable models, which Amazon CEO Andy Jassy flagged as potentially vulnerable to jailbreaking, became collateral damage in a dispute that administration insiders say reveals a fundamental culture clash. An administration official put it bluntly: "They screwed us."
The immediate trigger came when Jassy contacted Treasury Secretary Scott Bessent on Thursday with security worries about the models. But the real problem, according to multiple sources inside the administration, runs much deeper. Anthropic had already faced warnings weeks earlier after its Mythos model was made available to an entity with ties to the Chinese Communist Party. The company revoked that access without needing formal pressure, yet the overall pattern of engagement frustrated the administration.
What frustrated officials most was what they saw as Anthropic's refusal to take concerns seriously. When the Amazon report surfaced, "their position at the outset was no, we're not going to do anything, this is not a real issue," according to a source familiar with administration thinking. The company's response, a blog post dismissing the Amazon concerns, only made things worse. Anthropic then enlisted a cybersecurity expert the administration views as ideologically opposed to Trump's approach, who subsequently received praise from Chris Krebs, the official Trump just fired.
But beneath the specific disputes sits a communication problem that officials say Anthropic simply has not solved. "It's like they just speak in different languages," one source explained. "Anthropic has not done a great job at trying to speak to the administration and appreciate the ideological differences."
The situation echoes an earlier conflict between Anthropic and the Pentagon that also centered on personnel friction and messaging missteps. Even as the White House had been working to repair relations after that dispute, the latest breakdown happened quickly. One official described the situation as a case where "the technology is moving fast and the government is struggling to catch up," combined with personality clashes, resulting in a heavy-handed response instead of a calibrated one.
Anthropic counters that it received explicit government approval to deploy Fable and has consistently worked closely with regulators on these issues. A source close to the company denied it refused to address the jailbreak concerns.
The path forward remains unclear. The Commerce Department has scheduled meetings with Anthropic senior staffers Logan Graham, Dave Orr, and Nicholas Carlini for Monday, with additional sessions planned with the CIA and White House science advisor Michael Kratsios. Officials are considering whether Anthropic's models could be made jailbreak-resistant, though one source acknowledged that perfect resistance may be impossible.
An administration official suggested the real solution might be simpler: an attitude adjustment. "Instead of feeling dismissed," the official said, "everyone feels safe, secure and happy."
Author James Rodriguez: "This is what happens when a company built on the idea of slowing AI down collides with an administration obsessed with speed and dominance, all because nobody learned how to have an actual conversation."
Comments