When a Government Can Switch Off Your AI
Claude, export controls and the case for keeping local models within reach
On Tuesday, 9 June 2026, Anthropic released Claude Fable 5 and Claude Mythos 5. Three days later, neither model was available to anyone.
The reason was not a failed server, a cyberattack or a billing dispute. According to Anthropic, the United States government had applied immediate export controls to the new models. These controls required the company to restrict access for foreign nationals, both inside and outside the United States. Anthropic had no reliable way to verify every user’s nationality in real time, so it suspended access for everyone.
For most users, the result was simple: a powerful tool that existed on Tuesday had disappeared by Friday.
The restrictions were lifted on 30 June, and Fable 5 returned to general use the following day. Access to the less restricted and more capable Mythos 5 remained limited to approved organisations. The interruption was brief, but it revealed something much larger than a temporary product problem.
It showed just how little control we have over the artificial intelligence on which we are beginning to depend.
The cloud always has a landlord
Cloud services are wonderfully convenient. We do not have to buy expensive hardware, install large models or manage constant updates. We open a browser, enter a prompt and gain access to computing power that would have been unthinkable for an individual only a few years ago.
That convenience can create an illusion of ownership. We speak of “my Claude”, “my ChatGPT” or “my Copilot”, but the service is never truly ours. We are renting temporary access to somebody else’s infrastructure, under somebody else’s rules.
Several parties stand between the user and the model: the AI company, its cloud providers, payment processors, regulators and the government of the country in which it operates. Any one of them may change the conditions of access. An account can be restricted, a feature can be removed, a model can be retired, or an entire country can be excluded.
Until now, these possibilities often felt theoretical. The June intervention made them real. A government decision in one country temporarily affected users everywhere, including people who had no vote, representation or meaningful route of appeal in that country.
This matters because AI is quickly becoming more than an optional online service. Developers use it to write and review code. Researchers use it to analyse information. Businesses are placing it inside customer support, administration and security workflows. Individuals are using it for education, communication and creative work.
When a tool becomes infrastructure, the ability to withdraw access becomes a form of power.
Security concerns are real
It would be too easy to tell this story as one of government interference versus technological freedom. The actual situation is more complicated.
Anthropic said the government acted after researchers found a way to bypass some of Fable 5’s safeguards. The model could identify software vulnerabilities and, in one case, produce code demonstrating how a vulnerability might be exploited. Mythos 5 was deliberately built with fewer restrictions for a small group of trusted cybersecurity partners and reportedly possessed even stronger capabilities.
These are not imaginary risks. A sufficiently capable model could help defenders find weaknesses faster, but the same ability could also help criminals or hostile states attack vulnerable systems. Governments have a legitimate responsibility to consider those dangers.
Anthropic’s later testing, however, concluded that several older and more widely available models could reproduce the behaviour that triggered the intervention. The company deployed a new classifier that it said blocked the reported technique in more than 99 per cent of cases. It also called for a common framework that would distinguish a minor safeguard bypass from a genuinely dangerous jailbreak.
That distinction is important. If access to a major technology can be removed immediately, the criteria should be clear, consistent and open to scrutiny. Otherwise, safety becomes a word that can justify almost any intervention, while users and developers are left guessing where the boundaries lie.
The question is not whether powerful AI should have safeguards. It should. The question is who defines those safeguards, how decisions are reviewed and whether a small number of companies and governments should be able to make them for the rest of the world.
The beginning of an AI splinternet
The internet was once imagined as a borderless network. In practice, it has gradually fragmented into national and commercial zones. Content, services and privacy protections now vary according to location. Artificial intelligence appears to be following the same path much faster.
The most advanced models require enormous amounts of computing power. Only a handful of companies can build them, and most are based in the United States. This gives one jurisdiction extraordinary influence over which capabilities are available elsewhere.
Europe clearly sees the risk. On 3 June, shortly before the Claude restrictions, the European Commission announced a technological-sovereignty package covering semiconductors, cloud infrastructure, AI and open-source software. Its stated aim was to reduce strategic dependence and support European digital autonomy.
Technological sovereignty does not have to mean isolation. Europe should not attempt to disconnect itself from American innovation, just as individuals should not reject useful cloud services. Sovereignty means retaining meaningful alternatives. A system is not resilient if it stops working the moment a distant provider changes its terms or a foreign government changes its policy.
Local AI as a fallback, not a fantasy
This is where locally hosted and open-weight models become important.
A model running on a home server or workstation will not always match the newest frontier service. Hardware is expensive, energy use matters and installation still requires technical knowledge. Local models also need security updates and responsible configuration. They are not magically free from risk or influence.
But they offer one decisive advantage: continuity.
Once a suitable open-weight model has been downloaded, it cannot disappear from an online account overnight. It can continue summarising documents, searching a private knowledge base, translating text, assisting with code or handling other routine work without sending every request to a remote service. Sensitive data can remain on local hardware, and the user decides when the software is updated or replaced.
The sensible approach is not “cloud or local”. It is to avoid making the cloud the only option.
For individuals and small organisations, that could mean:
- keeping important documents and prompts in portable formats;
- testing a local model for essential, repeatable tasks;
- avoiding workflows that depend entirely on one vendor’s proprietary features;
- using open interfaces so models can be exchanged more easily;
- identifying which services would fail if an account or API suddenly became unavailable; and
- treating online AI access as a dependency rather than a possession.
The local model does not need to be the most intelligent system in the world. It needs to be capable enough to keep the lights on.
Who holds the switch?
I have previously written about the risk of AI becoming a new authority that people are expected to trust but cannot question. The events of June add another dimension to that concern. Authority does not only belong to the model that gives us an answer. It also belongs to whoever decides whether we may ask the question at all.
AI safety, export controls and national security will become more important as models grow more capable. Some restrictions will be justified. Others will be excessive, political or poorly designed. The only durable protection is a system in which power is distributed, rules are transparent and alternatives remain possible.
The temporary disappearance of Claude Fable 5 was not the end of open access to AI. Most users regained the model within weeks. But it was a useful warning.
The cloud may feel limitless, yet somewhere there is always a switch—and somebody else may have their hand on it.
Keeping a small model running locally will not solve every problem. It does, however, ensure that the roots remain alive when the canopy goes dark.
