US Government Forces Anthropic to Pull Top AI Models in Unprecedented Export Control Move

On June 12, 2026, the US Commerce Department's Bureau of Industry and Security (BIS) issued an extraordinary and unprecedented directive to the artificial intelligence company Anthropic. Citing unspecified national security authorities, the government ordered the immediate suspension of all access to the company's newly released Claude Fable 5 and Mythos 5 models for any foreign national, regardless of whether they were located inside or outside the United States. The sweeping mandate even applied to Anthropic's own foreign-national employees, creating an immediate operational crisis for the San Francisco-based firm.[1][2]

Because Anthropic operates a cloud-based API and web interface serving hundreds of millions of users globally, the company cannot reliably verify the nationality of every individual querying its servers in real time. Faced with the impossibility of selectively filtering out foreign nationals, the company was forced to execute a global 'kill switch.' The models were abruptly taken offline for all customers, including American citizens and enterprise clients, just days after their highly anticipated launch, sending shockwaves through the technology sector.[1][3]

This sudden intervention marks the first time the US government has successfully used export control authorities to restrict access to a deployed, cloud-hosted artificial intelligence model. Historically, the Export Administration Regulations (EAR) have been utilized to govern the transfer of physical, dual-use goods like advanced semiconductors, weapons systems, or discrete transfers of software source code and model weights. Applying these rules to a continuously available API service represents a profound shift in how the federal government views and regulates commercial software.[4][5]

Legal analysts and compliance experts note that applying the EAR to an API service represents a massive expansion of the Commerce Department's regulatory scope. By issuing this directive, the government is effectively treating the model's outputs—or the mere act of a user querying its servers from abroad—as a controlled export. This approach allows the Bureau of Industry and Security to bypass the traditional, lengthy notice-and-comment rulemaking process, enabling immediate enforcement but raising significant questions about due process and statutory limits.[4][6]

The administration cited urgent national security concerns to justify the sudden directive, pointing specifically to a potential 'jailbreak' that could theoretically bypass the models' built-in safety safeguards. Both Fable 5 and Mythos 5 possess highly advanced cybersecurity capabilities, including the ability to rapidly identify software vulnerabilities, analyze complex codebases, and potentially generate exploit code. Officials feared that if foreign adversaries gained unrestricted access to these capabilities, it could pose a severe threat to critical US infrastructure and digital networks.[1][7]

Anthropic strongly disputes the severity of the threat and the proportionality of the government's response. In a public statement, the company noted it received only 'verbal evidence of a potential narrow, non-universal jailbreak' rather than a catastrophic structural flaw. Anthropic's leadership argued that recalling a widely deployed commercial model over a minor, theoretical vulnerability sets a dangerous and unworkable standard. If applied consistently across the industry, they warned, such a precedent could effectively halt all future frontier AI development.[1][10]

The two models at the center of the controversy share the exact same underlying neural architecture but differ significantly in their behavioral constraints and intended audiences. Fable 5 is the consumer-facing version, equipped with strict refusals and conservative defaults designed to prevent the generation of harmful content. In contrast, Mythos 5 is an unsafeguarded version restricted to a small group of vetted partners, such as cyber defenders participating in the US government's Project Glasswing, allowing them to test offensive capabilities.[1][3]

Anthropic's own system card for the models acknowledges the inherent difficulties in securing advanced AI, stating plainly that 'perfect jailbreak resistance is not currently possible for any model provider.' Instead of relying on an impenetrable wall, the company utilizes a 'defense in depth' strategy. This approach combines secondary AI classifiers that actively monitor and block malicious prompts with a mandatory 30-day data retention policy, allowing security teams to continuously monitor for, analyze, and mitigate novel attack vectors as they emerge.[1][3]

The Trump administration views the situation through a fundamentally different lens. A senior administration official, speaking on the condition of anonymity, characterized Anthropic's response to the reported vulnerability as 'reckless.' The official claimed the company brushed off initial federal concerns and failed to take the jailbreak threat seriously enough, especially given the high expectations and security assurances set before the Fable 5 rollout. This perceived lack of cooperation reportedly eroded trust and directly precipitated the aggressive export control action.[8]

This unprecedented directive is not an isolated incident, but rather the latest escalation in a bitter, ongoing feud between Anthropic and the Pentagon. Earlier in 2026, President Trump ordered all federal agencies to immediately cease using Anthropic's technology. That sweeping ban was issued after the company strictly adhered to its terms of service and refused to allow its models to be integrated into fully autonomous weapons systems or used for domestic surveillance purposes by the US military and intelligence community.[7][10]

Following that high-profile refusal, Defense Secretary Pete Hegseth took the extraordinary step of designating Anthropic a 'supply chain risk'—a severe national security label historically reserved for foreign adversaries and state-backed telecommunications firms like Huawei and ZTE. Anthropic is currently fighting that designation in federal court, arguing it lacks any factual or legal basis. The export control directive is widely viewed by industry insiders as a continuation of this broader campaign to bring the defiant AI lab to heel.[2][10]

The unilateral US action has sparked widespread international alarm and diplomatic friction. Foreign researchers, multinational corporations, and allied governments suddenly lost access to a critical frontier model without warning. This abrupt disruption highlights the inherent vulnerability of relying entirely on American-controlled AI infrastructure. Consequently, the incident has dramatically accelerated debates over 'AI sovereignty' in Europe, the UK, and Asia, as foreign leaders realize the US possesses both the capability and the willingness to unilaterally cut off access to foundational technologies.[7][9]

For enterprise AI teams and corporate compliance officers, the directive introduces a completely new class of operational and legal risk. Companies that have spent months building complex workflows, customer service tools, and data analysis pipelines around specific frontier models must now account for the reality that the US government could force those models offline at a moment's notice. This paradigm shift is fundamentally altering how businesses evaluate cloud AI dependencies, pushing many to explore open-source alternatives or multi-model redundancy strategies.[2][5]

As the dust settles, it remains entirely unclear how long the global ban will last or what specific technical mitigations the Commerce Department will require before lifting the directive. The legal limits of the Bureau of Industry and Security's authority over intangible API access will almost certainly be tested in federal court. How the judiciary interprets these decades-old export laws in the context of cloud-hosted neural networks will ultimately set the rules of engagement for the next generation of artificial intelligence.[2][4]