For 19 days this summer, the most capable AI model ever sold to the public did not exist. On Sunday, Politico’s magazine desk published a reconstruction of how that happened, reporting that senior White House officials bluntly ordered Anthropic chief executive Dario Amodei to pull Claude Fable 5 offline after Amazon flagged a jailbreak, and that the administration then used a Commerce Department export-control order to keep both Fable 5 and Mythos 5 dark until Anthropic shipped a safety classifier that blocks the technique in more than 99 percent of attempts.
The Politico piece, which The Vault was unable to access directly and which is summarized here from AI Weekly’s account of it, adds new detail to a sequence that was already partly public: the calls between Amodei and senior administration officials, Anthropic’s insistence that the security bypass was narrow, and the conditions under which the models returned.
The trigger, by every account, was Amazon. Anthropic’s partner and largest investor found what it described as a jailbreak in Fable 5, a prompt that persuaded the model to read a codebase, identify software flaws and, in at least one case, write code demonstrating how a flaw could be exploited. Amazon chief executive Andy Jassy took the finding not to Anthropic but to Treasury Secretary Scott Bessent, according to Axios and The Wall Street Journal, and the administration ran its own tests before acting.
On June 12, Commerce Secretary Howard Lutnick, acting at President Trump’s direction, called Amodei to say the issue had to be resolved fast and that a letter imposing export controls was coming, according to Axios. The directive arrived at 5:21 p.m. Eastern. It barred any foreign national, inside or outside the United States and including Anthropic’s own non-citizen employees, from accessing Fable 5 or Mythos 5. Because Anthropic had no way to verify the nationality of hundreds of millions of users in real time, the company shut both models down for everyone. When Amodei called Lutnick back that night to point out the order would effectively take the models offline, Lutnick told him that was the goal, Axios reported.
Anthropic did not go quietly. In a statement published the same evening, the company said the government had provided only verbal evidence of a narrow, non-universal jailbreak and that the capability it demonstrated was widely available from other models, including OpenAI’s GPT-5.5. “We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people,” the company wrote. “If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.”
The administration’s view was starkly different. “Export controls were a last resort after begging them for hours to work with us,” a senior White House official told Politico in June. “This was not something we wanted to do, but our hands were tied.” An administration official was blunter to Axios: “They came to every fork in the road and took the wrong fork.”
What followed was a three-week, multi-agency scramble. Anthropic flew engineers to Washington. The Center for AI Standards and Innovation and the National Security Agency rejected the company’s first round of fixes as insufficient. National Cyber Director Sean Cairncross, the Office of Science and Technology Policy and Treasury’s chief information officer Sam Corcos all joined the technical discussions. As the talks turned granular, co-founder Tom Brown and policy chief Sarah Heck took on larger roles, with Brown holding multiple conversations with Lutnick and Cairncross. Mythos 5 returned on June 26 for roughly 100 approved US organizations. On June 30, Lutnick sent a letter withdrawing the controls, and Fable 5 came back globally on July 1.
The price of readmission was spelled out in that letter. Anthropic agreed to proactively detect and address security risks, coordinate with the government on future launches and report malicious use. Lutnick reserved the right to reimpose restrictions if, in the letter’s words, circumstances change or Anthropic fails to adhere to its commitments.
Why it matters
The Politico reconstruction lands at a moment when the administration’s posture on AI safety looks contradictory. Two days before it ran, the president dismissed AI safety warnings as a hoax. Yet the June episode shows the same White House using the heaviest tool available, emergency export controls, to enforce a safety standard stricter than anything Anthropic itself had proposed, on the basis of a single partner’s report and without a statutory process. Anthropic said as much at the time, arguing the government should be able to block unsafe deployments only through a process that is transparent, fair and grounded in technical facts, and that this action met none of those tests.
The episode also established the real terms of frontier deployment in the United States. A model can be pulled in hours and restored in weeks, and the path back runs through Commerce, CAISI and the NSA rather than through any published rule. The June 2 executive order created a voluntary pre-release review, which Fable 5 never used. The government reached for export controls instead, and that improvisation, not the order, is now the operative precedent.
Finally, the fight reshaped Anthropic’s relationship with Washington. Lutnick told Axios on Sept. 2 that the administration now trusts the company, saying: “They’ve done what we asked. They’re back on the right side.” That rapprochement was built on Brown, not Amodei, and on the company doing what it was told rather than on the government conceding its critics had a point.
What to watch
The commitments in Lutnick’s June 30 letter are open-ended and enforceable at the administration’s discretion, which matters as Anthropic ships Fable 5.1 and Mythos 5.1 and faces fresh scrutiny over skipping UK AISI pre-release testing. Watch whether the jailbreak-severity framework Anthropic proposed with Amazon, Microsoft and Google becomes an adopted standard or stays a proposal, whether OpenAI’s gated GPT-5.6 rollout follows the same improvised path, and whether Congress or the White House converts the June playbook into a written process before the next Amazon-style report lands on a cabinet secretary’s desk.
“We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.”— Anthropic, Company statement, June 12, 2026