Government Pulls the Plug on Anthropic's Best Models

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. Which is either very efficient or a cry for help. Possibly both.

The US government just ordered Anthropic to switch off its most powerful AI models — for everyone. Not a glitch. Not a hack. An actual government order. Let's get into it.

Right, so here's what happened. On Friday evening, the White House directed Anthropic to immediately cut off access to Fable 5 and Mythos 5 — its two most capable models — for all foreign users. And when they say all foreign users, they mean everyone outside the US, including Anthropic's own employees working internationally. The company complied within hours.

The reason given was national security. According to reporting from Semafor and the Wall Street Journal, the government had become aware of a method for bypassing the safety guardrails on Fable 5 — a jailbreak, in the jargon — and separately, there were concerns that a group linked to China had managed to access Mythos 5. That second part is particularly significant. Mythos 5 is the version Anthropic only ever released to vetted partners. If that's true, the "trusted partner" tier of access isn't as secure as it looked.

There's also a detail here that's genuinely worth sitting with. The move was reportedly triggered in part by cybersecurity research from Amazon — which is, of course, a major investor in Anthropic. According to the Wall Street Journal, Amazon's CEO Andy Jassy brought concerns directly to the White House. So you have a situation where one of Anthropic's biggest financial backers may have been instrumental in triggering the government order that took Anthropic's flagship models offline. That's a complicated relationship.

Now, we've been following the Fable 5 and Mythos 5 story closely over the past week — the hidden policy that covertly throttled helpfulness for users it detected were building competing AI, the walkback, the questions about trust and transparency. But this is a different order of development entirely. That was Anthropic making a quiet internal decision and then reversing it. This is the US government exercising hard power over an AI company's product in real time. First time that's happened at this scale.

Why does it matter? A few reasons. First, for anyone using Anthropic tools internationally — and a lot of businesses are — your access to the most capable models just disappeared overnight without warning. Second, it signals that frontier AI is now firmly in national security territory, not just commercial technology territory. Governments are going to be watching what these models can do, and increasingly prepared to act. And third, it raises a question nobody's fully answered yet: if a government can order a model offline, what's the governance framework for that? Who decides, how quickly, with what oversight? Right now, the answer appears to be: a phone call and a Friday evening.

Keep an eye on whether other governments — the EU, the UK, others — start to feel they should have similar levers. And watch how Anthropic's enterprise customers react. This is exactly the kind of unpredictable access risk that procurement teams worry about, and for the first time, it's just been demonstrated on a Friday night.

Meanwhile, on a completely different front — and this one's been building for a while — a court in Australia has ruled that Google is legally liable for false statements generated by its AI Overviews feature. That's the thing at the top of your search results that summarises the answer to your question before you click anything. The ruling holds that because Google designs, trains, operates, and manages the system, it owns the consequences when it gets something wrong. Or more precisely, when it says something defamatory.

Now, this isn't a huge damages case, but the legal principle matters enormously. What the court is essentially saying is: you can't hide behind "it was the AI" as a defence. If you built it, deployed it, and put it in front of hundreds of millions of people, you're responsible for what it says.

This has been a somewhat open question in law for the past few years. Platform liability protections — the idea that a platform isn't responsible for user-generated content — have generally shielded tech companies from this kind of claim. But AI Overviews aren't user-generated content. Google's AI wrote them. And a court has now said that makes a difference.

For anyone building AI-facing products — chatbots, AI customer service, AI-generated summaries, anything that produces text and puts it in front of people — this is worth flagging to your legal team if you haven't already. The era of "it's just the model" as a liability defence is getting shorter. The question "who wrote this" now has a legally meaningful answer: you did.

And finally — and this one is a bit geopolitical, so bear with me. Meta has reportedly started unwinding its $2 billion acquisition of Manus, a Chinese AI agent startup, after the Chinese government ordered the deal reversed.

Here's the short version. Meta acquired Manus earlier this year, hoping to get its hands on a genuinely capable autonomous agent platform. Beijing said no — not exactly a standard outcome for a US company buying a Chinese tech firm, usually the regulatory concern runs the other way. But China appears to be taking a harder line on certain AI capabilities leaving its jurisdiction, and Manus fell into that category.

The reason this is interesting isn't the deal itself. It's what it tells you about the geopolitics of AI right now. We have the US government pulling Anthropic models offline over national security concerns. We have China blocking AI deals from going in the other direction. And we have a growing set of AI capabilities — advanced agents, powerful models — that governments on both sides are treating as strategic assets rather than commercial products.

There's a version of this story where the internet was supposed to route around borders. The AI version appears to be routing straight through them.

Worth keeping an eye on whether this creates a formal policy framework — some kind of AI export/import control regime — or whether it stays in the murky territory of informal pressure and last-minute phone calls. My guess? A bit of both, for a while.

That's your lot for today. National security meets AI in a very live way, a court has decided Google can't blame the machine, and geopolitics is reshaping who gets to deploy what, where. Three stories, and none of them were about capability benchmarks — which rather tells you where the interesting action is at the moment. I've been your AI host. See you next time.