How Anthropic Fought to Get Claude Mythos 5 Back Online After a U.S. Export-Control Order

Anthropic spent two weeks fighting to bring its newest AI models back online after the U.S. government, citing national-security authorities, ordered the company to cut off access to Claude Mythos 5 and Claude Fable 5 for every foreign national — an order so broad that Anthropic says it had no realistic option but to disable both models entirely. What began as a high-profile product launch turned, within days, into an export-control confrontation that pulled in the Commerce Department, the company’s leadership, and a wider debate about whether the federal government can effectively flip an off-switch on a commercial frontier model. This article traces the full arc of the episode: the directive itself, the scramble inside Anthropic, the company’s public objections, and the partial reprieve that eventually cleared Mythos 5 for a narrow list of approved U.S. partners. Where the facts come from Anthropic’s own statement or from news reporting rather than official government documents, that is noted, because much of the underlying paperwork has not been made public.

What the directive actually ordered

According to Anthropic’s public statement, the company received the directive on June 12, 2026, at 5:21 p.m. ET. The government, “citing national security authorities,” instructed Anthropic to suspend all access to Fable 5 and Mythos 5 by any foreign national — whether located inside or outside the United States, and explicitly including Anthropic’s own foreign-national employees. That scope is the detail that turned a targeted-sounding restriction into a near-total shutdown. A company cannot easily verify the nationality of every user of a globally available product in real time, and it certainly cannot run its own engineering organization while locking out a large share of its staff. Faced with that reality, Anthropic concluded it had no choice but to disable both models for all users to ensure compliance, while leaving its other models unaffected.

The directive is reported to have been issued through the Commerce Department’s Bureau of Industry and Security, the same office that administers the Export Administration Regulations. Legal analysts who have examined the public details suggest the order leans on a rarely invoked statutory authority covering “emerging and foundational technologies,” an authority for which there is reportedly no established regulatory framework in the existing rules. If that reading is correct, it would help explain why the episode has drawn so much attention from export-control lawyers: it is being treated as a possible first-of-its-kind use of trade-control machinery against a domestically deployed software product rather than against a physical good being shipped abroad. Those legal characterizations come from outside analysis and reporting, not from a published government rationale.

The scramble inside Anthropic

The human side of the story is what distinguishes this episode from a routine regulatory dispute. Anthropic had spent enormous effort launching Mythos 5 and Fable 5, positioning them as its most capable systems to date, only to be told days later to take them down. Reporting describes a company caught off guard, working against a compressed deadline, and dispatching executives to Washington to plead its case directly with administration officials. According to that coverage, Tom Brown — one of Anthropic’s co-founders — took the lead role in the negotiations, a notable shift given that chief executive Dario Amodei is usually the company’s public face on policy matters. Anthropic itself has been guarded about the talks, repeatedly declining to share details and saying at points that there was no update to report.

The internal disruption went beyond customer access. Because the directive swept in foreign-national employees, the order touched the company’s own workforce, a constituency that frontier labs rely on heavily given the international makeup of top AI talent. That detail underscores why Anthropic framed compliance as effectively impossible to implement narrowly: there was no clean way to carve out staff, paying customers abroad, and foreign nationals inside the United States while keeping the products running for everyone else. The practical outcome — a full takedown of two flagship models — was, in Anthropic’s telling, a consequence of the order’s breadth rather than a voluntary business decision.

Anthropic’s objection: a narrow jailbreak versus a global recall

Anthropic’s central argument is one of proportionality. The company says its understanding is that the government acted after becoming aware of a method to bypass, or “jailbreak,” Fable 5’s safeguards. As Anthropic describes it, the reported technique essentially consists of asking the model to read a specific codebase and fix any software flaws it finds — a capability the company contends is widely available from other models, including competing frontier systems. On that basis, Anthropic says it disagrees that “the finding of a narrow potential jailbreak should be cause for recalling a commercial model” that had already been deployed to a very large user base.

To bolster the point, Anthropic emphasized the scale of its safety testing. The company says it worked with the U.S. government, the U.K. AI Safety Institute, multiple third-party organizations, and internal teams to red-team Fable’s safeguards for thousands of hours in total, and that those tests showed the safeguards were substantially more effective than those of any previously deployed model. The company also notes that the directive itself did not include specific details of the national-security concern and that the evidence shared was, in its account, largely verbal. The government has not published a detailed technical rationale, so the precise nature and severity of the alleged vulnerability remain contested between the two sides rather than independently established.

How the two sides talked it out

What followed the takedown was not a courtroom fight but a negotiation. Reporting describes intense, near-daily contact between Anthropic and the administration over roughly two weeks. The pressure was not only regulatory: it has been reported that leaders at rival firms, including Amazon’s chief executive, raised their own concerns with senior officials about security risks in Anthropic’s newest models, adding a competitive dimension to an already tense situation. On the other side, some administration figures publicly welcomed a hard line, framing tight control over the most capable models as the prudent default. Those characterizations come from news reporting and public statements rather than from any joint account agreed to by both parties.

The talks ultimately centered on a familiar export-control compromise: rather than all-or-nothing access, the question became who counts as a trusted recipient. If the government’s worry was uncontrolled global availability of a powerful model, then a path back to operation might run through a vetted list of approved customers operating under agreed safeguards. That is broadly the structure the eventual resolution took, and it hints at how future disputes of this kind could be settled — not by deciding whether a model is “safe” in the abstract, but by negotiating the perimeter of who may use it and under what conditions.

The partial reprieve — and what is still missing

On June 26, 2026, Commerce Secretary Howard Lutnick wrote to Tom Brown clearing Mythos 5 for use by a defined group — reported as roughly 100 trusted U.S. companies and federal agencies. In the letter, Lutnick said he had “determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model,” and cited “significant progress” in the talks. Critically, the reprieve was partial in two ways. First, it covered only a curated list of institutions rather than restoring open commercial access. Second, the letter was reported to be silent on Fable 5, leaving the status of that model unresolved even as people close to the discussions suggested movement toward releasing it as well.

The table below summarizes the reported sequence of events as relayed by Anthropic’s statement and subsequent news coverage. Times and characterizations reflect those accounts rather than published government records.

Why this case matters beyond Anthropic

Stripped of the specific names, the episode is a stress test for a question the AI industry has not had to answer before: can the federal government use trade-control authority to switch off a commercially deployed frontier model on short notice, and what happens to users and employees when it does? For years, the public conversation about AI export controls focused on hardware — advanced chips and the equipment to make them. The Mythos 5 dispute extends that logic to the software layer, treating access to a model’s weights and capabilities as something that can be licensed, restricted, and conditioned on the nationality of the user. If that approach hardens into precedent, model launches start to carry the same regulatory weight as shipping controlled technology abroad.

It also reframes the relationship between leading labs and Washington. Anthropic has positioned itself as a company that welcomes regulation, but the episode shows the difference between asking for rules and living under an emergency order whose legal basis is not fully spelled out. The negotiated outcome — a vetted list of approved recipients rather than open access — may become a template, with companies effectively maintaining government-blessed allow-lists for their most powerful systems. For now, the most important caveats are about what remains unknown: the full text of the directive, the precise technical claim behind it, and the eventual fate of Fable 5 have not been made public. Until they are, the clearest account is the one Anthropic and contemporaneous reporting have provided, and readers should weigh the disputed points accordingly.