When Washington Pulls the Plug: The Anthropic Cutoff and the New AI Sovereignty Scramble
A sudden US export order blocking access to two cybersecurity models has accelerated Europe's push to build its own AI stack - and exposed the fragility of cross-border technology dependence.
An Overnight Blackout
On a Friday in mid-June, Anthropic found itself in an unusual position: the company was ordered by Washington to shut off access to two of its cybersecurity-focused AI models, Mythos 5 and Fable 5, for anyone who was not a US citizen. The directive was so sweeping that even some of the company's own employees could no longer use the tools they had helped build. Because there is no reliable way to verify citizenship over the internet, Anthropic effectively had to pull the models offline for everyone.
The company itself appeared caught off guard. According to Anthropic, the US government cited concerns about a potential jailbreak - a method to bypass the model's safety guardrails - though the evidence provided was described as verbal and limited in scope. The specific vulnerability, as Anthropic understood it, involved prompting the model to analyze a codebase and identify software flaws. Whether that constitutes a genuine security risk or an overly cautious interpretation of capability remains unclear. At the time of writing, both models remain unavailable.
At DailyTechWire, we have tracked export controls on semiconductor tooling and chip architectures for years, but the extension of such measures to software models - especially with minimal notice and ambiguous justification - marks a different kind of intervention. It also raises a question that policymakers in Brussels, London, and beyond are now asking with renewed urgency: what happens when the infrastructure you rely on can be switched off by a foreign government?
The Brussels Response
The European Commission wasted little time in framing the incident as a case study in dependency risk. Thomas Regnier, a Commission spokesperson, said the body is assessing the implications of the directive for European users and noted that such measures should not discriminate against partners. He emphasized that the episode illustrates why Europe must strengthen its technological sovereignty, pointing to existing frameworks like the AI Act, the Cyber Resilience Act, and the NIS2 Directive as tools to manage such risks independently.
The timing is notable. Just days before the Anthropic order, the Commission had launched its European Technological Sovereignty Package, a suite of initiatives designed to reduce reliance on US and Chinese technology across multiple domains. The package includes funding for domestic chip fabrication, cloud infrastructure, and AI model development. The Anthropic incident, from Brussels' perspective, is not an anomaly - it is a preview of what dependency looks like when geopolitical priorities shift.
Europe's flagship AI effort, Mistral AI, has gained traction as a fast, open-source alternative to frontier models from Anthropic and OpenAI. But even supporters acknowledge that Mistral does not yet match the performance of the most advanced US systems. The trade-off is becoming clearer: accept a performance gap in exchange for operational control, or remain tethered to platforms that can be revoked without consultation.
The Invisible Cost of Dependency
Aled Lloyd Owen, chief of staff at Responsible AI UK, argued that the incident strengthens the case for accelerating European independence in AI. He framed the challenge as both technological and commercial: building effective models and infrastructure is one part of the equation, but weaning European companies off high-capability overseas systems is another. The question, he suggested, is whether organizations are willing to sacrifice perceived or real performance for reliability and sovereignty.
Kate Hanaghan, chief research officer at TechMarketView, put it more bluntly. She recalled recent conversations with European systems integrators who described dependency risk as a cost that stays invisible until it materializes. The Anthropic order, she noted, made that risk visible. For enterprises in the UK and across Europe, relying on a single US frontier provider now carries a clear operational hazard: access can be withdrawn without warning, and contingency plans are rarely in place.
The comment that resonated most widely came from Kanishka Narayan, the UK minister for AI and online safety, who posted on social media that access to AI capabilities is now a matter of national security. He argued that the UK treats other sovereignty threats with seriousness but has not yet learned to do the same for AI. There is no simple switch to flip, he acknowledged, but the central political question of the era is whether Britain will build the capability it needs or allow someone else to decide the answer.
The Jailbreak Question
The substance of the US government's concern remains murky. Anthropic has stated that it was given only verbal evidence of a narrow jailbreak involving the model's ability to read and analyze code for vulnerabilities. The company has long maintained that perfect jailbreak resistance is not feasible with current techniques and that no universal jailbreak for its models has been demonstrated. It continues to advocate for a defense-in-depth approach: layered safeguards rather than a single point of failure.
Anthropic also noted that the features cited as problematic in Mythos 5 and Fable 5 are present in other models, including OpenAI's GPT-5.5. Whether those systems will face similar restrictions is an open question. The inconsistency - or the perception of it - has fueled criticism that the order was reactive rather than strategic, driven more by a specific conversation than by a coherent risk framework.
A group of fifty-four security and AI researchers responded with an open letter calling for the directive to be lifted and for future decisions of this kind to be grounded in transparent, democratic processes. The signatories did not all agree on whether the US government should have regulatory authority over AI releases, but they converged on the view that materially impactful decisions should involve security teams, give advance notice, and restrict capabilities only to the extent necessary. The letter argued that the order removed the best tools from defenders, created market uncertainty, and risked American AI leadership without a proportional security justification.
Cybersecurity Models and the Dual-Use Dilemma
Cybersecurity-focused AI models occupy a particularly sensitive niche. They are designed to identify vulnerabilities, analyze attack vectors, and assist in defensive operations - but the same capabilities can, in theory, be repurposed for offensive use. This dual-use nature is not new; it has been a feature of security research for decades. What is new is the speed and scale at which these models can be deployed, and the difficulty of controlling access once they are released.
Mythos 5 and Fable 5 are still early-stage systems. They are not yet widely adopted by enterprises, let alone casual users. But their trajectory is important. If these models prove effective, they will become embedded in security operations, threat intelligence workflows, and vulnerability management pipelines. At that point, the ability to revoke access becomes a form of leverage - one that allies and adversaries alike will notice.
The Anthropic case also highlights the limitations of export controls designed for hardware. Semiconductor manufacturing equipment can be tracked, inspected, and physically restricted. Software models, by contrast, can be copied, fine-tuned, and redistributed with relative ease. Enforcing export controls on AI models requires either technical measures - such as authentication and usage monitoring - or the kind of blanket shutdown that Anthropic implemented. Neither is ideal. The former is complex and can be circumvented; the latter is indiscriminate and disruptive.
What Europe Can Realistically Build
The European Technological Sovereignty Package is ambitious, but it is also constrained by the realities of capital, talent, and time. Building a competitive AI stack requires more than funding. It requires access to compute infrastructure at scale, which in turn depends on semiconductor supply chains that Europe does not fully control. It requires attracting and retaining researchers who might otherwise join better-funded labs in Silicon Valley or Beijing. And it requires a policy environment that can move faster than the technology it seeks to govern.
Mistral AI is a start, but it is not a complete solution. The model is open-source, which aligns with European preferences for transparency and interoperability, but it lags behind frontier systems in certain benchmarks. Closing that gap will take years, and there is no guarantee that European models will catch up before the next generation of US or Chinese systems arrives.
There is also the question of infrastructure. Training large models requires access to advanced GPUs and tensor processing units, most of which are designed and manufactured by American companies like Nvidia or fabbed in Taiwan. Europe's efforts to build domestic chip capacity, such as the European Chips Act, are progressing, but they will not yield results overnight. In the meantime, European AI labs remain dependent on the same supply chains that underpin their US competitors.
The Sovereignty Trade-Off
Sovereignty in AI is not a binary condition. It exists on a spectrum, and the choices that governments and enterprises make will determine where they land. At one end is full dependency: relying on foreign providers for models, infrastructure, and updates, with all the convenience and risk that entails. At the other end is full autonomy: building every layer of the stack domestically, which offers control but at significant cost and with no guarantee of performance parity.
Most European organizations will end up somewhere in the middle. They will use a mix of domestic and imported models, hedge their bets with multi-cloud strategies, and invest in contingency planning. But the Anthropic incident has clarified the stakes. The cost of dependency is not just financial or operational - it is strategic. When access can be revoked overnight, the systems you build on top become fragile.
For governments, the calculus is different. National security, economic competitiveness, and regulatory authority all point in the direction of greater sovereignty. But achieving it requires sustained investment, cross-border coordination, and a willingness to accept trade-offs in the short term. The European Commission's rhetoric is strong, but the real test will be whether member states align on funding, procurement, and industrial policy.
The Road Ahead
The Anthropic order is unlikely to be the last of its kind. As AI models become more capable and more embedded in critical infrastructure, the pressure to control their distribution will intensify. Export controls, licensing regimes, and usage restrictions will all be part of the policy toolkit. The challenge will be designing those measures in a way that manages genuine risks without fragmenting the global research ecosystem or creating unnecessary friction for allies.
For Europe, the path forward involves building capacity while managing expectations. Mistral and other domestic efforts will continue to improve, but they will not replace frontier US models in the near term. In the meantime, European organizations will need to diversify their dependencies, invest in fallback options, and prepare for a world where access to cutting-edge AI is conditional rather than guaranteed.
At DailyTechWire, we have followed the region's policy responses to semiconductor export controls, cloud sovereignty debates, and data localization mandates. The AI sovereignty push is the next chapter in that story, and the Anthropic case has given it new momentum. Whether that momentum translates into durable infrastructure and competitive models remains to be seen. But the urgency is no longer in question.
