Anthropic has been forced to suspend two of its artificial intelligence (AI) models globally after the US government issued a directive prohibiting their use by foreign nationals over “national security” concerns.
The move to suspend Anthropic’s models came just days after the launch of the Claude Mythos 5 frontier model, with the US government claiming it had “become aware” of a jailbreaking method that allowed its safeguards to be bypassed.
Access to Mythos 5 has been restricted by the firm and is only available through a “trusted access program” of around 50 organisations – including cyber security partners and researchers – that use it to conduct advanced cyber security analysis to find and patch vulnerabilities before the technology is available to attackers.
Fable 5 is the consumer-facing version built with extra safeguards, which the US government found a method of jailbreaking.
Anthropic made clear in its statement that it did not agree with the US government’s decision to recall the AI models. It pushed back on claims of the models’ vulnerability, describing it as “minor” and “relatively simple”, adding that “if this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers”.
This is the first time the US government has exercised an export control directive against an AI firm. In response, more than 80 cyber security leaders, including Nvidia and Adobe, signed an open letter in support of Anthropic, urging the government to lift the control directive and ensure regulations are “enforced transparently and fairly”.
Growing concerns over UK’s reliance on US tech
The episode between Anthropic and the US government comes amid increasing concern among UK MPs about the over-reliance on US tech companies, which has previously led to calls for the UK government to back sovereign IT.
Following the suspension, Baroness Beeban Kidron questioned the government in the House of Lords about whether the increased dependence on US companies in health, education and security created a critical vulnerability for national security.
Many of today’s technological dependencies were established in an era of relative geopolitical stability and a baseline assumption of trust. Recent events, however, demonstrate that this assumption has been undermined Sharinee Jagtiani, German Marshall Fund
Responding to Kidron on behalf of the government, Baroness Lloyd of Effra did not directly address the concerns raised about national security, but did stress the need for the UK to develop sovereign AI capabilities. “Our approach is about building strength over key parts of the value chain to bring to the table technologies that no one else can do without,” she said.
Concerns over the UK’s reliance on US technology were also raised by MPs in April 2026, after a report by the Open Rights Group (ORG) flagged the national security and economic risks to the UK’s critical infrastructure, and the potential of services being shut down by US order.
ORG previously warned, in January 2026, that hyperscale cloud services such as Microsoft Azure and Amazon Web Services (AWS) are bound by US law, meaning American authorities hold the power to demand access to data stored in the UK.
Sharinee Jagtiani, a geopolitical analyst at the US-based German Marshall Fund, said: “Many of today’s technological dependencies were established in an era of relative geopolitical stability and a baseline assumption of trust. Recent events, however, demonstrate that this assumption has been undermined.”
The next step should be focused on mitigating risk and shifting towards “managed interdependence” to increase resilience, said Jagtiani. European countries are likely to start by advancing policy initiatives to support domestic technology ecosystems and reduce exposure to external shocks.
“Global businesses and public institutions must now operate in an environment where technological dependencies can be weaponised, and factor that in as they build out their business models and policies,” Jagtiani added.
This includes strengthening a country’s role in emerging technology chains, developing a diverse supplier base rather than relying too heavily on a single partner – for example, by deepening connections with “technology middle powers” such as Canada, Australia, Japan, South Korea and India.
Anthropic versus the US government
Tensions between Anthropic and the US government have been escalating for some time. This began when Anthropic’s CEO, Dario Amodei, denied the US military use of its AI models for “fully autonomous lethal weapons or mass domestic surveillance of Americans” in February 2026.
The US government responded by putting Anthropic on a national security blacklist – with defence secretary Pete Hegseth labelling the company a “supply chain risk” – meaning the service is not secure enough for government use. Anthropic sued the US Department of Defence over this move.
Claude was prohibited from use in the Pentagon as a result. Despite this, the AI model was used by the US military in the conflict in Iran, as reported by CNBC.
A US federal judge ultimately ruled that the Pentagon’s directive could not be enforced, finding that the government’s measures appeared designed to punish Anthropic rather than address any genuine security concern.
Commercial organisations
This incident has similarly sparked conversation within commercial organisations that rely on US-owned AI models.
Mary Mesaglio, analyst at business management consultancy Gartner, said this incident emphasised the “need for organisations to be intentional about sovereignty dependencies and, where possible, to design model-agnostic architectures”.
She added that CIOs and AI leaders must stabilise AI-dependent workflows now and use this event to focus on model concentration risk, talent-technology resilience and sovereign AI disruption.
From a cyber security perspective, Jamie Moles, a senior technical manager at US-based AI cyber security company ExtraHop, said: “The concern over a ‘jailbreak’ that uncovers software flaws highlights why a defence-in-depth strategy is non-negotiable.”
He added: “Perfect model resistance doesn’t exist, and it’s important to never rely on just one model, as an action like this can create significant setbacks for the individual user and wider projects across an organisation.”
