If you use Anthropic’s newest AI models in your business, you had a rough week. The Trump administration hit Anthropic with an export control order that forced the company to shut down access to its Fable 5 and Mythos 5 models for everyone, not just overseas users. Foreign nationals inside the US, including Anthropic’s own employees, were caught in the sweep. Nobody got a clear explanation why.
This isn’t a theoretical policy debate anymore. AI export controls just became an operational problem for any business that depends on frontier models. And the rules are so vague that even the experts can’t tell you what they actually require.
What happened
- The Trump administration ordered Anthropic to cut access to Fable 5 and Mythos 5 for all foreign nationals, including users and employees inside the United States.
- Anthropic had to block access to both models for everyone in order to comply.
- The government cited “national security authorities” and referenced an alleged jailbreak linked to groups in China. Anthropic pushed back, saying the jailbreak did not allow users to bypass all of its safeguards.
- The administration has not publicly explained the legal basis for the order.
- According to Georgetown’s Center for Security and Emerging Technology, this is believed to be the first time US export controls have been used to restrict access to an AI model in this way.
- Export controls have traditionally applied to things you can ship, copy, or download (weapons, hardware, software, source code). A cloud-hosted chatbot doesn’t fit neatly into that framework.
- Remote access to cloud services is a known gap in current export control rules, and Congress is already working on legislation to close it.
The numbers
- 0 public explanations from the administration about the legal basis for the order.
- First known case of US export controls being applied to restrict access to an AI model, per Georgetown researchers.
- 2 models (Fable 5 and Mythos 5) pulled offline, affecting every user, not just foreign ones.
5 things every service business operator should know about AI export controls
- Your AI tools can disappear overnight. Anthropic didn’t choose to shut down its models. The government forced it to. If you’re building workflows around frontier models, you now have a new category of risk that has nothing to do with uptime or pricing.
- The rules don’t make sense yet. Export controls were designed for physical goods and downloadable software. A chatbot running on someone else’s servers doesn’t fit the existing framework. Experts at Georgetown and UC Berkeley both called this area “unsettled” and “unsustainable.” You can’t comply with rules nobody can define.
- This probably won’t stop at Anthropic. If these models were targeted because they’re uniquely capable, the same logic applies to the next powerful model from OpenAI, Google, Meta, or xAI. If they were targeted because of safeguard concerns, every lab is exposed. Either way, expect more of this.
- Foreign team members are a compliance issue now. The order applied to foreign nationals inside the US, not just people overseas. If your team includes anyone who isn’t a US citizen, their access to certain AI tools could be restricted without warning.
- International clients add another layer. If you serve clients outside the US, or if your clients have international teams, the tools you build on top of frontier models could suddenly become inaccessible to some of those users. That’s a contractual and operational problem, not just a policy one.
The hot take
The US government is actively sabotaging its own AI lead. You can’t champion American AI dominance while yanking frontier models offline through secret orders with no legal transparency. UC Berkeley professor Andrew Reddie put it plainly: if the standard becomes that models must be impossible to jailbreak, “then [the US] will have no AI models.” This kind of ad hoc enforcement doesn’t protect national security. It just teaches every other country that American AI platforms are unreliable. That’s the fastest way to push global customers toward alternatives built in Europe, the Middle East, or China.
The Agency OS play
This week, do an honest audit of your AI dependencies. Write down every frontier model your business relies on, every workflow that touches one, and every team member or client who accesses them. Flag anyone on your team who is a foreign national. Flag every client with international operations. You need to know your exposure before the next order drops, not after.
Then start building redundancy. If you’re running everything through one model provider, you’re one government order away from a full stop. Set up fallback workflows that can switch to a different model (or a different provider entirely) with minimal disruption. Test them now, while things are calm. The businesses that weather this kind of disruption are the ones that treated it as inevitable.
Finally, if you serve clients in regulated industries or across borders, add a compliance checkpoint to your onboarding process. Ask where your client’s team members are located. Ask about citizenship. It feels awkward, but it’s better than discovering mid-project that half your client’s team can’t access the tools you built for them. Document everything. When the rules finally get clarified (and they will), you’ll want a paper trail showing you took this seriously.