Washington put up two gates in front of the most powerful AI models this summer. On 2 June 2026 President Donald Trump signed an order that lets developers offer frontier models to national security agencies for up to 30 days of testing before release. Ten days later the government used a harder tool: export controls that took Anthropic's newest models, Claude Fable 5 and Claude Mythos 5, offline for every user in the world. What it showed users outside the United States: access to a model can be cut off by nationality. Both episodes are the background to the voluntary accord of 29 September, covered in our lead story on the Super Intelligence order.

The June order: a 30 day look before release

Executive Order 14409, "Promoting Advanced Artificial Intelligence Innovation and Security", sets up voluntary national security vetting of the most advanced AI systems for up to 30 days before public release, the Associated Press reported. The Treasury Department, the National Security Agency (NSA) and the Cybersecurity and Infrastructure Security Agency (CISA) must first set benchmarks for what counts as a frontier model, according to Roll Call, while AP reports that the government, led by the NSA director, decides which models qualify. The Congressional Research Service (CRS) dates the deadline for that classified benchmarking process and the review framework to 1 August 2026, 60 days after signing.

Nothing in the order authorises mandatory licensing, permitting or preclearance for AI models, the law firm Holland & Knight notes. The White House said it was "NOT conducting oversight of all new models". Trump had declined to sign an earlier version on 21 May: "We're leading China, we're leading everybody, and I don't want to do anything that's going to get in the way of that lead."

Reactions: a first step, or not enough

Juan Londoño of the Cato Institute called the order imperfect but "a step in the right direction". He also told AP that so much discretion for the NSA director was a "dangerous precedent" that could let the government "weaponize" the policy against companies it is clashing with, such as Anthropic.

Others want reviews to be compulsory. Republican Senator Josh Hawley backed the order but said: "I would go farther." He wants his bill with Senator Richard Blumenthal to make "that sort of reporting and monitoring mandatory." Brendan Steinhauser, chief executive of the Alliance for Secure AI, said: "Congress must now codify the White House's EO with legislative action."

Chris Lehane, OpenAI's chief global affairs officer, said safety frameworks "should continue to be developed through democratic institutions". Democratic Senator Mark Warner, vice chairman of the Senate Intelligence Committee, welcomed the policy but said the administration had "belatedly discovered the need to redo something it hastily dismantled in its first year."

12 June: export controls on two models

Anthropic released Fable 5 and Mythos 5 on 9 June. They share one underlying model, the company says: Fable 5 shipped with strong safeguards for general use, Mythos 5 went only to a few partners for defensive cybersecurity. On 12 June the government applied export controls to both. Anthropic had to cut off every foreign national, inside or outside the United States, including its own non citizen staff, The Hacker News reported.

The trigger was a report by Amazon researchers who got Fable 5 around its safeguards: it identified software vulnerabilities and, in one case, wrote code showing how one could be exploited. Anthropic said weaker models, such as GPT-5.5 and Kimi K2.7, could identify the same flaws. "We disagree that the finding of a narrow potential jailbreak should be cause of recalling a commercial model deployed to hundreds of millions of people," the company said, as quoted by IAPP.

"Export controls were a last resort after begging them for hours to work with us," a senior White House official told Politico, according to IAPP. David Sacks, the former White House AI adviser, backed the move, saying the issue, "while serious, should be easily resolved. The ball is in Anthropic's court." Nearly 80 cybersecurity executives and experts asked for the controls to be lifted, arguing the action had "taken the best models away from defenders".

How access came back

Anthropic trained an improved safety classifier aimed at the reported technique; blocked requests now go to the weaker Opus 4.8, at the cost of more false alarms in routine coding. The Commerce Department's Center for AI Standards and Innovation tested the safeguards, the company says.

More than 99%
of cases using the technique from the Amazon report are blocked by Anthropic's new classifier, the company said on 30 June 2026.

On 26 June Mythos 5 came back for roughly 100 US companies and federal agencies that defend critical infrastructure, The Hacker News reported. On 30 June the controls were lifted. "Over the past two weeks, we have worked closely with Anthropic to analyze and approve Fable 5," Commerce Secretary Howard Lutnick wrote on X, as CNN reported. Fable 5 returned worldwide on 1 July. In return, Anthropic agreed to coordinate on future launches and report malicious use.

What it means outside the United States

The Cloud Security Alliance called the directive "the first publicly documented exercise of export control authority over a commercially deployed frontier AI model". The tool, an "Is Informed" letter from the Bureau of Industry and Security, had until then been used for chip exports. Because Anthropic "had no reliable way to verify nationality in real-time", the models went dark for all users, including customers on AWS, Google Cloud and Microsoft Foundry.

The models went dark globally because nationality could not be verified in real time, so a control aimed at foreign nationals became an outage for everyone.
Sanchit Vir Gogia, chief analyst at Greyhound Research, to CIO, 1 July 2026

"This week the most advanced AI model on the planet got switched off by a foreign government," said Al Carns, a British MP and former armed forces minister, quoted by IAPP. OpenAI, too, first released its GPT-5.6 models only to "a small group of trusted partners" at the government's request, CIO reported.

From gate to accord

Fable 5 never went through the voluntary review; the government used export controls instead, The Hacker News observed. On 29 September the White House added a third, voluntary layer: an accord, signed by executives of Meta, Nvidia, Google, OpenAI, xAI and Anthropic, that calls for internal controls, an internal oversight team, outside auditors and an independent board committee, with no penalties and no duty to publish audit results, Al Jazeera reported. More in our lead story.