US export controls force Anthropic to disable Fable 5 and Mythos 5

The directive that landed in Anthropic's inbox at 5:21pm ET on June 12 was not a chip restriction or a future-licensing rule. It ordered the company to suspend all access to Fable 5 and Mythos 5 by any foreign national — whether sitting in London, Lagos, or a desk in San Francisco — including Anthropic's own foreign-national employees ^[1]. Commerce Secretary Howard Lutnick invoked national security authorities under the Export Control Reform Act of 2018, the first time the department has aimed that power at an AI model, and warned Dario Amodei that non-compliance would bring criminal and civil penalties ^[2]. Faced with the impossibility of cleanly gating a model used by customers worldwide by checking each user's citizenship, Anthropic chose the blunt path: it switched both models off globally for everyone, while leaving other Claude models running ^[1]. Fable 5 was Anthropic's first general release in the Mythos class, shipping with classifiers to block high-risk responses like cybersecurity; Mythos 5, distributed to a vetted consortium under Project Glasswing, runs with some of those constraints removed and carries genuine dual-use reach — identifying software vulnerabilities and developing exploits ^[1][11]. The government's stated worry was that Fable 5's guardrails could be bypassed to reach Mythos 5-class capability. Anthropic disputes the framing, calling the trigger a narrow, non-universal jailbreak that amounts to asking the model to read a codebase and fix flaws — hardly grounds, it argues, for yanking a product used worldwide ^[1].

The most uncomfortable detail is who set the machinery in motion. Amazon's cybersecurity team used a series of prompts to coax the Mythos-class model into producing restricted cyberattack information, and CEO Andy Jassy carried the finding straight to senior administration officials; by Friday evening the Commerce Department had issued the controls ^[3]. Amazon is not a bystander here — it is one of Anthropic's largest investors and the cloud host that runs its models, making it investor, landlord, and whistleblower at once. The second irony cuts deeper. Days earlier, Dario Amodei had been publicly calling for binding AI regulation, including the power to block unsafe models from deployment; Washington obliged, then went much further than he intended, and Anthropic now labels the result a 'misunderstanding' it is racing to undo ^[8]. The capabilities at the center of the dispute are not hypothetical: a Project Glasswing launch partner that had been testing Mythos since April reported the model surfacing 26 CVEs in a single scan, against a typical monthly volume of fewer than five — the kind of leap that collapses attack timelines from weeks to minutes. Yet critics counter that pulling one vendor's model solves little, since comparable capability is arriving across other labs and open-weight releases regardless ^[11].

For enterprise buyers, the episode crystallized a risk that had been abstract until June 12. Deutsche Bank strategist Jim Reid put it plainly: 'You can't rely on something that could be switched off' ^[12]. The Peterson Institute's Martin Chorzempa drew the same line toward open models — when you cannot know whether the US government will cut off access to future releases, weights you can own, inspect, and audit start to look less like a compliance headache and more like insurance ^[12]. The market moved accordingly. Mistral's Arthur Mensch and DeepSeek pitched sovereign alternatives, Chinese open-source firms Zhipu and MiniMax saw their stocks surge, and France's intelligence agency moved to replace Palantir's AI tools with a domestic provider ^[4][12]. The political reaction split sharply: former Trump-administration AI hand Dean Ball dismissed the order as 'simply cartoonish,' while ex-AI czar David Sacks defended it as serious-but-solvable, putting 'the ball in Anthropic's court.' Roughly 80 cybersecurity executives and experts signed an open letter urging the White House to lift the Fable 5 restrictions and commit to an open, scientific, transparent process for future AI risk assessments — a rare instance of the security community lining up against a security-justified ban. Public reaction split along the same fault line: free-speech advocates on X framed a citizenship-based cutoff as a civil-liberties problem, while European commentators pushed back that an Americans-only policy 'should not be discriminatory.'

Conventional wisdom said a government shutdown of your flagship product would bleed customers and investors. The data says otherwise — at least so far. Anthropic surpassed OpenAI in business AI spending share in May 2026, reaching 41% according to Ramp, with usage of its Opus models (especially Opus 4.8) climbing, and the company had already posted its first profitable quarter before the government acted ^[5]. That resilience sits in tension with the financial overhang the directive created: Anthropic raised $65 billion at a $965 billion valuation and filed confidential IPO paperwork, and the demonstration that Washington can darken its frontier products overnight cuts directly against the predictability investors price in ^[5][9]. The contradiction is the story — adoption momentum and regulatory tail risk pulling in opposite directions on the same balance sheet.

In a February 2026 clash, rivals had rallied to Anthropic's defense. This time the sympathy thinned. OpenAI and others offered less support, with some arguing Anthropic invited the scrutiny by simultaneously talking up its models' dangers and demanding aggressive regulation ^[7]. The talent dimension worried OpenAI most directly: CSO Jason Kwon reminded staff that US AI leadership leans on global talent — by one MacroPolo estimate cited in the reporting, 38% of researchers publishing at leading AI conferences in 2024 received their undergraduate education in China — and OpenAI began auditing its own export exposure ^[6]. The wreckage even reached the negotiating table: Trump advisers have been weighing whether to take equity stakes in AI firms, with Treasury Secretary Scott Bessent favoring using such stakes to seed Trump Accounts and Lutnick preferring a sovereign wealth fund — and the export controls on Anthropic's models have complicated those government-equity talks ^[10]. Anthropic's senior technical staff are in Washington trying to reframe the whole affair as a misunderstanding; whether they succeed will set the template for how the US treats every frontier model that follows.