The AI race, interrupted

A version of this article originally appeared in Quartz’s AI & Tech newsletter. Sign up here to get the latest AI & tech news, analysis and insights straight to your inbox.

The most powerful AI in the world was available for a few days. Maybe you got to try it. Most people didn't, and now, thanks to the U.S. government, nobody can.

Earlier this month, the Trump administration issued an export control order forcing Anthropic to take its two most powerful AI models, Fable 5 and Mythos 5, offline for all users. The official reason was a potential security vulnerability that could allow someone to bypass the models' safety guardrails. 

So now American companies, researchers, and regular subscribers have all lost access to what many observers described as a significant advance in AI capability, even as Anthropic scrambles to smooth over the latest drama in an already dramatic year for Anthropic.

What actually happened

The sequence of events is still disputed. Researchers at Amazon $AMZN, which is both an investor in and a major computing partner to Anthropic, reportedly alerted the Trump administration to a potential jailbreak of Fable 5. The White House gave Anthropic roughly 90 minutes to take the model down. 

When the company didn't comply fast enough, officials issued an export control, a trade restriction that bars foreign nationals from accessing the technology. Because so many Anthropic employees, like those at most frontier AI labs, are not U.S. citizens, the company had little choice but to pull the models for everyone.

Anthropic disputed the White House’s reasoning, saying the actions the jailbreak potentials were either entirely benign or minor vulnerabilities that other publicly available models could identify just as easily.

Cybersecurity experts who reviewed the reported jailbreak were underwhelmed. Independent analysts found the vulnerability amounted to a security researcher asking Fable to fix deliberately insecure code, which is more or less what the model was designed to do. OpenAI's GPT-5.5, which faces no such restrictions, has similar capabilities. 

More than 100 cybersecurity professionals from universities including MIT and companies like Zoom $ZM signed an open letter arguing that the export control removed the best tools from defenders without meaningfully reducing risk.

Anthropic did not exactly help itself here. The company spent months describing Mythos as too dangerous to release, then put out a version of it with guardrails two months later. That strategy, meant to signal responsibility, may have instead handed critics a ready-made argument. 

"One reaps what one sows," AI researcher Yann LeCun wrote on social media after the export control was announced. The company insisted the guardrails were robust but having loudly declared your own technology capable of serious harm, it becomes harder to argue later that the government is overreacting to a security concern.

Tails I win, heads you lose

The government's response, however, has its own credibility problem. While Anthropic's models are offline, the DOJ is actively defending xAI in court against a lawsuit over unpermitted gas turbines at its Mississippi data center, arguing that Grok, which has never been considered a top-tier model, is vital to military operations.

That favoritism is also hard to square with the administration's own recent policy moves. At the beginning of June, Trump signed an executive order asking AI companies to voluntarily submit new models for government review before release, a framework Anthropic publicly supported. 

The administration scrapped an earlier version of that same order, partly over concerns it would slow development and hurt competitiveness with China. Now it has done what that order was designed to avoid, pulling a model from the market with no clear process, no timeline, and no public accounting of the evidence.

The White House has tangled with Anthropic before. In February, the Pentagon labeled the company a supply chain risk after a dispute over a Defense Department contract, a designation that legal experts said lacks a clear legal basis and may be illegal. That didn’t stop Defense Secretary Pete Hegseth from celebrating the export control on social media this month. Meanwhile, the NSA has reportedly sought exceptions to use Mythos for its own cyber operations.

Senior Anthropic staff is in Washington meeting with White House officials, and both sides said they want to resolve the dispute. Whether they can may depend less on the technical details of the jailbreak and more on the political dynamic between a safety-focused AI lab that has never been shy about disagreeing with the government and an administration that has shown little tolerance for that posture.