Anthropic's most advanced AI models disappeared from the internet for three weeks this summer following a dramatic clash between the company, the Trump administration, and competing claims about a critical technical vulnerability. The dust has settled, but the episode exposed deep fractures in how the U.S. government and Silicon Valley handle AI safety crises.
The standoff began in early June when Amazon, a major investor in Anthropic, flagged concerns about a "jailbreaking" flaw in the company's latest models, Mythos and Fable. The issue involved a breakdown in the AI systems' safety guardrails. Amazon escalated the warning to federal officials, triggering swift action from Commerce Secretary Howard Lutnick, who moved on President Trump's direction to impose sweeping export controls that effectively forced Anthropic to take the models offline.
What happened next revealed how quickly disagreement could fracture the response. Within days, leading cybersecurity experts sent an open letter to the administration disputing Amazon's framing, arguing that similar vulnerabilities existed across rival AI models from OpenAI, Google, and others. The implication was clear: singling out Anthropic looked arbitrary.
Lutnick's June 12 call to Anthropic CEO Dario Amodei left no room for interpretation. The Commerce Secretary made clear that the company had to act fast. When Amodei called back after reviewing the export control letter, Lutnick confirmed that yes, the models needed to come down. It was a coordinated, high-pressure play.
What followed was unusual for the Trump administration: a coordinated, multi-agency technical sprint involving the National Security Agency, the Commerce Department, Treasury, the newly created Center for AI Standards and Innovation, the White House Office of Science and Technology Policy, and other officials. Anthropic deployed engineers to Washington to work directly with government specialists. The company's initial position, according to officials, was that the jailbreaking problem was already solved and only minor tweaks remained. Federal agencies disagreed and demanded deeper fixes.
Treasury Secretary Scott Bessent emerged as a quiet but crucial broker in the talks. Bessent was the first recipient of Amazon's alarm and helped navigate White House Chief of Staff Susie Wiles and other officials back toward engagement with Anthropic. He also represented the administration's position internationally at G7 meetings, underscoring that the U.S. was serious about global AI safety standards.
Behind the closed-door meetings, frustration mounted. Personality clashes and poor communication threatened to derail the entire negotiation. The turning point came when Anthropic realized that confronting the government would fail; they needed to align with it instead. As discussions shifted into technical minutiae, co-founder Tom Brown took a more prominent role alongside CEO Amodei. Brown's technical background allowed him to sit with government engineers and debate how models behaved under stress, line by line through code.
On July 1, after nearly three weeks offline, Anthropic's models returned to the market with federal approval for their safety measures. But the resolution masked unresolved tensions. When OpenAI sought to release its latest model, GPT-5.6, it found itself in separate technical discussions with the same agencies. Unlike Anthropic, OpenAI had not been included in earlier talks and operated with limited visibility into what the government expected. The company entered negotiations essentially blind.
Officials now face a larger problem. Neither Anthropic nor other AI labs know when or how they will be permitted to distribute advanced models to allied nations, a step many believe is necessary to compete with China's AI development. There is no clear process, no published timeline, and no transparency standards for how future models will be approved.
The three-week shutdown revealed a government willing to use export controls as a pressure tool and an industry that had no playbook for managing federal intervention in real time. Both sides made it up as they went. Whether that ad hoc approach will work for the next crisis, or the one after that, remains genuinely uncertain.
Author James Rodriguez: "This wasn't just about fixing code, it was about who controls the narrative around AI risk, and the government made sure everyone knew it was calling the shots."
Comments