Friday, June 12, 2026, 5:21 p.m. ET.
Anthropic receives a letter from U.S. Secretary of Commerce Howard Lutnick. The directive: immediately suspend all access to Claude Fable 5 and Mythos 5 for foreign users. Both models had been launched barely three days earlier. Within hours, they were offline worldwide, for everyone, including European companies that relied on them daily.
Last week, this became a reality.
Why this also affects European companies
What at first glance appears to be a political conflict between the Trump administration and a Silicon Valley company touches on a fundamental risk for any organization that uses third-party AI models. This is the first time a leading tech company has been forced to take its most powerful AI models completely offline. It may not be the last.
The underlying tensions between Anthropic and Washington have been simmering for some time. In February 2026, the Pentagon labeled Anthropic a supply chain risk after the company refused to budge on two key principles: no mass surveillance of U.S. citizens and no fully autonomous weapons without human oversight. The immediate cause of last week’s shutdown lay elsewhere. According to Anthropic, the secretary’s letter did not cite any specific security reason. The company inferred from verbal communications that the government was concerned about a “jailbreak” technique, in which the model analyzes a codebase to detect vulnerabilities. According to the Wall Street Journal, Amazon reported that vulnerability directly to the U.S. Department of Commerce.
Anthropic itself calls the measure a misunderstanding. The company states that the same capabilities are also available in other publicly available models, including GPT-5.5, and is working to restore access.
The order specifically applied to foreign users, including Anthropic’s foreign employees on U.S. soil. Anthropic was unable to filter its users by nationality in real time. A global shutdown was therefore the only way to comply with the order. Any European company relying on these models was immediately left out in the cold—without warning.
The crux of the problem: architectural dependency
At Xylos, we see this pattern more often than one might think. Organizations are building powerful AI applications—such as copilots, automation, knowledge management, and customer interaction—all built on top of a single external model. Speed of implementation takes precedence, and architectural risk is often overlooked.
But what if that platform is gone tomorrow? The reason might have nothing to do with you. A geopolitical conflict, a government decision, a takeover, or a security incident is all it takes to suddenly cut off access.
This is what we call model dependency risk, and we believe it deserves a prominent place in every AI strategy.
How does Xylos handle this?
We always advise our clients based on the same principle: AI architecture must be vendor-agnostic and resilient by design.
In practical terms, this means:
- Multi-model strategy. Make sure your applications aren’t hardcoded to a single provider. Abstraction layers such as LLM orchestration frameworks allow you to quickly switch between models from Anthropic, OpenAI, Google, Mistral, and other providers.
- On-premises and open-source alternatives. Local models such as Llama, Mistral, or Phi can be a valuable addition for specific use cases. They offer control over data, independence from external APIs, and protection against sudden outages like this. For a well-defined task, they are often sufficient, and they significantly reduce your vulnerability. You won’t always get the absolute best performance from them, which isn’t a problem for many use cases.
- Data and prompt sovereignty. Your prompts, fine-tuning data, and knowledge bases are strategic assets. Keep them separate from the model provider so that, if you switch, you can quickly rebuild on a different foundation.
- Business Continuity Planning for AI. Just as you plan for disaster recovery for your infrastructure, you need an AI continuity plan. What is your fallback if Model X becomes unavailable tomorrow? Who makes the decision? How quickly can you switch over?
Being aware is the first step
We remain convinced of the value of AI-as-a-service. The power of models like Claude and GPT is real, and the benefits they deliver are significant. Every powerful tool brings with it a degree of dependency, and that dependency deserves to be managed thoughtfully.
Recent events show that the battle over who controls the future of autonomous intelligence is in full swing. Today, geopolitics and technology are inextricably linked to regulation. As an organization, you cannot influence that battle. But you can anticipate it.
Xylos helps companies build AI architectures that perform today and can withstand the uncertainties of tomorrow. An architecture that is robust and agile—and, above all, remains yours.
Do you have questions about how to make your AI strategy more resilient? We’d be happy to discuss it with you.
About the author
