Anthropic lifts export curbs as AI sovereignty fears grow

The US government has lifted export controls on Anthropic's Mythos and Fable artificial intelligence models, prompting a wider debate among vendors and policymakers over the resilience and security of AI infrastructure.

The initial restriction, which limited foreign access to some of Anthropic's most advanced systems, exposed how quickly geopolitical or regulatory decisions can disrupt the supply of closed foundation models. AI founders and government technology advisers argue the episode will have lasting effects on how states and enterprises structure their AI stacks, even as access resumes.

Eugene Cheah, Chief Executive Officer and Co-Founder of Featherless.ai, said the episode had already shifted sentiment toward open-source alternatives that organisations can deploy on their own infrastructure.

"This is welcome news from Anthropic. However, I would be concerned that the damage has already been done. It showcased how fragile access to closed models really is and may push more governments, companies, and developers toward open models they can run, adapt, and keep access to themselves.

"That is the point of open source: research is not limited to a small number of labs, and teams can tune models for the problems they are trying to solve. Rio 3.5 Open, a model released by the City of Rio de Janeiro, is a good example. It shows how far open-model work can be pushed when it is adapted for a specific use case," he said.

Featherless.ai, which recently closed a USD $20 million Series A round led by AMD Ventures and Airbus Ventures, focuses on infrastructure that allows enterprises to run specialist models across varied hardware in their own jurisdictions. Cheah said this approach reduces reliance on a small group of hyperscale providers and gives organisations greater control over where and how AI systems operate.

He expects closed models to retain an edge on the most complex workloads, but said most businesses do not need premium-grade systems for every task.

"Closed models will still be ahead in some areas, especially on the hardest and most expensive tasks. But most organisations do not need a Porsche for every journey. They need models that are good enough for the job, that they can keep access to, and that are not dependent on decisions made by a handful of companies," Cheah said.

Anthropic has also begun redeploying Claude Fable 5, a version of its Mythos tool that it had previously judged too risky for broad release. The system includes new safeguards and user limits, and Anthropic has acknowledged that "releasing a model this capable comes with risks".

Roman Stanek, Founder and Chief Executive Officer of GoodData.AI, said longstanding failures in basic cybersecurity remain the bigger issue around such advanced models.

"The thing with AI and cybersecurity is that the vulnerabilities we're worried about AI exploiting, well, we've known about most of them for 20 years. We just never fixed them. Not because we couldn't, but because nobody wanted to pay for them. Open-source security, legacy code debt, infrastructure hygiene - they are all solvable problems, but all chronically underfunded. Now we have Mythos, this newer, shinier, more expensive model that can theoretically find and fix them.

"Great. Except the incentive structure hasn't changed. Nobody wanted to pay a human engineer to fix it. They're not going to pay an AI to fix it either. The issue has never been capability. People are just unwilling to invest, and until that changes, the vulnerabilities will remain as the threat evolves. We will continue to have the same conversation around better and better tools that nobody will use," Stanek said.

For public authorities, the incident has intensified discussion of "AI sovereignty". John Harms, Head of Government Solutions at Quantexa, said it underlined the importance of control over the data and decision-making layers that sit above any particular model.

"Government IT leaders regularly tell me that uncertainty is a significant concern this year. Recent events surrounding access to Anthropic's Mythos and Fable models illustrate why. Even though the decision affecting foreign access has been reversed, the episode serves as a reminder that organisations can be exposed to risks outside their direct control.

"This is not a US-versus-the-rest-of-the-world issue. Any government or organisation, regardless of geography, needs confidence that the technologies underpinning critical operations will remain available. The lesson is clear: resilience depends on sovereignty. And sovereignty requires 'control under distress'," Harms said.

He defined that as the ability to sustain operations through geopolitical, regulatory, commercial, or technical disruption while retaining authority over data, models, and the software that turns outputs into real-world decisions.

"Control under distress is the ability to continue operating confidently during geopolitical, regulatory, commercial, or operational disruption. It means being able to govern data, AI models, decision-making processes, infrastructure, and operational dependencies under all circumstances," Harms said.

He said governments do not need to insist on purely domestic models, but must avoid structures in which external "kill switches" can halt critical services.

"This doesn't mean we should reject all global technology and insist that every model is built in the UK. That would limit choice and weaken competition. But organisations need to be in a position where they still control their data and decision intelligence layer for critical aspects of the tech stack.

"The prospect of externally controlled 'kill switches' raises understandable concerns. Public sector leaders should look at every level of their technology stack, from infrastructure and models all the way to the software on which decisions are being made," Harms said.