Anthropic pulls top AI models after US government order

The US government ordered Anthropic to switch off its most powerful models

On Friday 12 June 2026, Anthropic disabled Claude Fable 5 and Mythos 5 for every customer worldwide. The US government had ordered it to suspend access for all foreign nationals.

The two models had launched only three days earlier. The directive arrived at 5:21pm Eastern time. Within hours, two of the most capable models on the market went dark for everyone.

What the order actually says

Commerce Secretary Howard Lutnick sent the directive to Anthropic chief executive Dario Amodei. It cited national security and required a licence to export or transfer the two models.

The restriction covers any foreign national, inside or outside the United States. It even covers Anthropic's own foreign-national staff. The company was given no detailed explanation of the concern.

Anthropic says the government believes there is a jailbreak. In theory it could let Fable 5 help identify software vulnerabilities. The company calls the evidence verbal and the jailbreak narrow and non-universal.

Why Anthropic switched everything off

Anthropic cannot screen foreign nationals out of its user base in real time. So it chose the blunt option. It turned the models off for everyone rather than attempt a partial block.

This appears to be a first for the industry. No leading AI provider has pulled a live, public model offline at a government's order before. Access to all other Anthropic models continues as normal.

Anthropic disagrees with the decision. It argues that recalling a model used by hundreds of millions over a narrow jailbreak is disproportionate. It says rival models can find similar bugs in code.

Its sharpest line is a warning to the whole industry. If this standard applied to everyone, it said, all new frontier model deployments would effectively stop.

Expect more of this, not less

This is the part worth sitting with. As models get more capable, governments will take more interest in them. The same skill that finds software bugs for a developer can find them for an attacker.

That dual-use problem does not go away. It gets sharper with every release. We expect more interventions, more export controls, and more sudden changes to what is available at the frontier.

For years, export controls focused on chips and the hardware behind AI. This order targets access to the model itself. That is a meaningful shift in how governments treat this technology. Treat it as the new normal, not a one-off.

The political backdrop is messy

The timing invites suspicion. Anthropic's relationship with the government ruptured this year. The company refused to let the US military use its models for domestic surveillance and autonomous weapons.

The Department of Defense then labelled Anthropic a supply chain risk. Anthropic sued over the designation. The dispute had recently shown signs of easing, with the company preparing for a public listing.

Some will read this order as the government settling a score. Others will take the security concern at face value. From the outside, it is genuinely hard to say which is true.

Our advice is not to over-read the politics. Whatever the motive, the outcome is the same. A model that businesses might have started using vanished overnight.

Safety over speed is the right instinct

Some people will moan about heavy-handed regulators slowing progress. We see it differently. Caution at the frontier of a powerful technology is sensible, not timid.

Anthropic itself recently proposed that leading AI firms coordinate to pause advanced development. It has argued the technology is improving fast enough to warrant real care. A government acting cautiously sits comfortably with that.

You can disagree with this specific order and still back the principle. Getting safety right matters more than shipping the newest model a week sooner.

What this means for your business

Here is the practical lesson for any UK SME. Do not build your critical work on the single newest, most powerful model. That model is the one most likely to be pulled, restricted, or changed.

The good news is that you rarely need it. The models from six months ago handle the vast majority of everyday business tasks. Drafting, summarising, classifying, and answering questions do not need the absolute frontier.

Resilience comes from your approach, not from chasing the top of the leaderboard. AI adoption is 80% people and culture, 20% technology. The model is the part you should worry about least.

Three things you can do this week

• Check which model your AI tools actually run on. Know what you depend on before it changes.
• Stay platform-agnostic. If your work only runs on one provider, a single order like this can stop you.
• Write your core processes down as clear instructions. Good Core Documents and prompts move between models. The work survives even when a model does not.

Next step

Frontier models will keep coming and going. The businesses that cope best are built on steady habits and clear processes. They have people who know how to use the tools they already have. gecco helps UK SMEs build exactly that, on whichever platform suits you. Take the AI readiness assessment to see where your team stands.

Sources

Reporting from Reuters, NBC News, Al Jazeera and Tom's Hardware, plus Anthropic's published statement of 12 June 2026.