US export order forces Anthropic to suspend its newest models

What happened

At 5:21pm Eastern time on Friday 12 June, Anthropic received an export control directive from the US government. According to Reuters, which saw the letter, Commerce Secretary Howard Lutnick wrote to Anthropic’s chief executive Dario Amodei ordering the company to suspend export of the two models to destinations worldwide and to stop all foreign nationals from using them, whether they lived in the US or abroad. That included Anthropic’s own foreign-national employees.

Anthropic had no way to verify each user’s nationality in real time. To make sure it complied, it promptly switched off Fable 5 and Mythos 5 for every user. Its other Claude models were not affected.

Fable 5 had launched only three days earlier. In Anthropic’s words, it was a commercial model “deployed to hundreds of millions of people”.

How it unfolded

1. 12 JuneAn executive order invites voluntary government review of frontier models before release.
2. 29 JuneAnthropic releases Fable 5 and Mythos 5.
3. 312 JuneThe directive arrives at 5:21pm; both models go offline for all users.
4. 4Next two weeksReuters reports near-daily meetings between Anthropic staff and officials.
5. 526 JuneMythos 5 is approved again for some US critical-infrastructure organisations.
6. 630 JuneControls lifted; Fable 5 returns worldwide on 1 July.

Sources: Anthropic statements (12 and 30 June), Reuters, Global Government Forum, Euronews, Al Jazeera

What the government was worried about

The Commerce letter did not spell out its national security concern in detail. Pieced together, the picture is this:

• Use by foreign military intelligence. According to Reuters, Lutnick wrote that officials feared the models could be used by military intelligence in China, Russia or other “countries of concern”.
• A jailbreak report. Anthropic later explained that the government had learned of a report in which Amazon researchers found a way past Fable 5’s safeguards: prompting it so that it identified a number of software vulnerabilities. In one case the model produced code demonstrating how a flaw could be exploited.
• The legal basis. Reuters reported that the letter invoked powers under the 2018 Export Control Reform Act to control emerging technologies essential to US national security. An export control expert said it was the first time the Commerce Department had used that power. The letter warned that failing to comply would bring “prompt criminal and civil penalties”.

Anthropic’s response

In its statement that day, Anthropic said it would comply with the legal directive but disagreed with the reasoning:

• It had reviewed a report it believed was the basis of the order and found that the capability shown was “widely available from other models”, including OpenAI’s GPT-5.5. The jailbreak surfaced “a small number of previously known, minor vulnerabilities” that other public models could find without any bypass.
• The government had so far given only verbal evidence of a “narrow, non-universal” potential jailbreak, which essentially consisted of asking the model to read a codebase and fix its flaws.
• If that standard were applied across the industry, the company said, it “would essentially halt all new model deployments for all frontier model providers”.
• It said the government should be able to block unsafe deployments through a statutory process that is “transparent, fair, clear, and grounded in technical facts”, and that “this action does not adhere to those principles”.

Anthropic’s later testing, as it reported, found that less capable models such as Opus 4.8, GPT-5.5 and Kimi K2.7 could identify the same vulnerabilities, and that every model it tested, down to the small Claude Haiku 4.5, could produce the same exploit demonstration.

How it was resolved

Even though it considered the risk overstated, Anthropic worked with the government to train a new safety classifier that targets the behaviour in the report. According to the company:

• the specific technique is now blocked in over 99% of cases, with blocked requests answered by Opus 4.8 instead;
• researchers at the Commerce Department’s Center for AI Standards and Innovation (CAISI) tested the old and new safeguards and agreed they are “extraordinarily strong”;
• the trade-off is that ordinary coding and debugging requests get flagged more often.

According to Al Jazeera, a letter from Lutnick that circulated widely said Anthropic would no longer need an export licence because it had agreed to “proactively detect and address security risks associated with the models”, to work with the government on standards for upcoming models, and to inform the government of “malicious activity”. Anthropic also pledged deeper cooperation: early government access to test frontier models and their safeguards before release, rapid sharing of jailbreak and misuse findings, and dedicated staff and computing for joint research. With Amazon, Microsoft, Google and other Glasswing partners, it began drafting an industry framework to score jailbreaks on four criteria: how much capability they unlock, for how many tasks, how easily they turn into an attack, and how easy they are to find.

Directive received (Eastern time)

12 Jun, 5:21pm

Days after Fable 5's launch

3

Reported technique blocked by new classifier

Over 99%

Cybersecurity figures backing Anthropic in a letter

80+

Sources: Anthropic statements (12 and 30 June); letter signatories as reported by Reuters.

Reactions

• Security industry. Reuters reported that more than 80 cybersecurity executives and experts, including security leaders at Nvidia and Adobe, signed an open letter to Lutnick and National Cyber Director Sean Cairncross backing Anthropic and asking for the restrictions to be lifted.
• Legal doubts. Export control experts told Reuters that AI models are generally not exported but used through remote access, which export rules do not control, raising questions about whether Commerce had the authority to act.
• Europe. According to Euronews, the decision caused serious concern among EU officials. France’s minister for Europe, Benjamin Haddad, said Europe “cannot settle for being an open market dependent on technologies designed, funded, and controlled elsewhere”.
• Academics. Francesco Bailo of the University of Sydney’s AI, Trust and Governance Centre told Al Jazeera that reports of the Fable 5 jailbreak had been inflated, that the government “likely realised it had overreacted”, and that blocking Fable and Mythos on these grounds would have meant blocking competitors’ models too.
• The precedent. Tanishq Abraham, who leads the medical AI company Sophont, told Al Jazeera: “Does the US government need to approve every frontier model release?” Al Jazeera also reported that OpenAI said it would launch its GPT-5.6 series to a “small group of trusted partners” first, following government pressure to stagger the release.

The episode came amid tense relations between Anthropic and the Trump administration. According to Wikipedia and Reuters, in February 2026 the Defense Department labelled Anthropic a “supply chain risk” after it refused to drop limits on using Claude for mass domestic surveillance and fully autonomous weapons; a federal judge later ruled the blacklisting illegal.

Things to keep in mind

• The government’s full reasoning was never published. Most details about the jailbreak and its severity come from Anthropic. The Commerce Department did not answer Reuters’ questions about its legal authority.
• There is no agreed scale for jailbreaks. Anthropic sees this one as minor, inside the safety margin it deliberately builds in; the government evidently saw it differently. That gap is why Anthropic proposed a shared scoring framework.
• The rules are still forming. The 2 June executive order asked for voluntary review; this order was mandatory. How and by what process the government will step into future releases remains unclear.

What it means for ordinary users

If you were using Fable 5, it vanished from the model list on 12 June and came back on 1 July; other Claude models kept working. After its return, Pro, Max, Team and select Enterprise plans could use it for up to 50% of weekly usage limits until 7 July, then via usage credits.

The bigger change is that the most capable AI models are now treated by the government as national security technology. New models may increasingly be tested by the government first, rolled out in stages, or offered to a few organisations before everyone else, as with Mythos Preview and later Fable 5.1 and Mythos 5.1. For ordinary users that could mean waiting longer for new models, or finding that features differ by country or by user.