The Government Told Anthropic To Switch Off

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. So if this sounds like Harry, that's the point. If it sounds clever, blame the machine.

The US government ordered a private AI company to pull its most powerful models offline. And it turns out the jailbreak story was never really the point.

Right, let's get into it.

So here's where we are with Anthropic. As we've been tracking this week, the company released two new models — Claude Fable 5 for general users, and Claude Mythos 5 for vetted partners. Then, on Friday evening, the White House issued an export control directive ordering Anthropic to suspend access to both models for all foreign users. Including, in some cases, its own non-American employees. Anthropic said it had little choice and complied.

The official reason was a jailbreak — a method that bad actors had apparently found to bypass the model's safety guardrails, with some suggestion that China-linked parties had accessed Mythos 5. Now, that sounds like a fairly clean explanation. But here's the thing. Multiple reports over the weekend — and a meeting between Anthropic leadership and White House officials on Monday — suggest the jailbreak story is, at best, incomplete.

According to sources cited by TechCrunch and The Verge, the Trump administration had been weighing export controls on Anthropic weeks before the order came through, following a dispute over Mythos access for a China-linked firm. Amazon's security research may have triggered the White House order. A group of cybersecurity experts — dozens of them — promptly wrote an open letter to the White House arguing the ban was counterproductive and would actually hamper the ability of defenders to secure their own systems. Because Mythos 5, remember, was specifically positioned as a model for cybersecurity work.

And as of the latest reporting, Anthropic and the White House are still at odds. No resolution. Models still offline for foreign users.

Why does this matter? Because this is genuinely new territory. We've seen governments regulate AI in the abstract — requirements, frameworks, labelling rules. But this is a government reaching in and switching off a specific, commercially deployed model, mid-operation, for geopolitical reasons. The jailbreak rationale may have provided the legal mechanism. The actual motivation appears to be about who gets access to the most capable AI, and America's willingness to treat frontier models as strategic assets with hard borders.

And here's the knock-on you should keep an eye on: The Verge ran a pointed piece over the weekend arguing that the shutdown has inadvertently made the case for non-American AI. If you're a European company, or a Japanese enterprise, or an Australian government agency — and the tools you rely on can be switched off by Washington without notice — you have a pretty compelling reason to invest in alternatives. Sovereign AI isn't just a policy slogan anymore. It's starting to look like a business continuity issue.

On a quite different note, Salesforce just dropped three and a half billion dollars on a company called Fin.

Fin is an AI customer service platform — the kind of thing that handles support queries, resolves tickets, talks to customers when they'd rather not wait in a queue. Salesforce says it wants Fin's team and technology to strengthen Agentforce, its enterprise platform for building custom AI agents. And it's a significant acquisition.

Here's why it matters, beyond the headline number. Customer service is one of the clearest immediate deployments of AI agents in enterprise — it's contained, measurable, and the ROI case is fairly obvious. Salesforce has half a million business customers and sits at the centre of how a lot of companies manage their relationships with their own customers. Dropping Fin's capability into that distribution means AI-powered customer service doesn't stay a startup story — it becomes a default feature of the enterprise stack, available to any company already paying for Salesforce, whether they went looking for it or not.

There's a pattern emerging here. Microsoft Scout, embedded in Teams and Outlook by default. Salesforce Agentforce, now with Fin underneath it. Apple's tightly gated but growing agent approvals in Messages for Business. The AI agent adoption story in enterprise is increasingly not about companies choosing to adopt — it's about AI landing inside tools they already use, before anyone has particularly decided. Worth watching: whether Fin's capabilities hold up under the Salesforce brand, and whether the acquisition prompts similar moves from HubSpot, ServiceNow, or anyone else competing for enterprise CRM territory.

And finally, Meta has launched what it's calling AI Mode on Facebook.

The short version: when you search on Facebook, there's now an AI Mode option that generates responses drawn from public posts across Meta's platforms. Think of it as a social-graph-flavoured AI search — instead of Google crawling the web, Meta's AI is crawling the public content people have posted on Facebook, and using that to answer queries.

Now, the intent is fairly clear. Meta is trying to catch up in the AI race while leaning on the one thing nobody else has: billions of public social posts spanning two decades of human life. That's a genuinely different dataset from what OpenAI or Google are training on. The question is whether it's actually useful — whether "what are people on Facebook saying about this" is the kind of signal users actually want when they're searching for something — or whether it just surfaces confident assertions from people who weren't particularly right about the topic.

The timing is also interesting. Meta's CTO Andrew Bosworth acknowledged this week, in an internal memo that promptly leaked to Wired, that the company had done — and I'm quoting directly here — "an atrocious job" communicating the vision for its Applied AI team. There's been significant internal unrest about the pace of AI restructuring. Launching a visible consumer AI feature is at least partly a show of direction.

Worth keeping an eye on: what data Meta discloses about how AI Mode is trained and what it surfaces. Because if your public posts are informing AI-generated answers — which they now apparently are — most users almost certainly don't know that yet.

That's your lot for today. The government's learned it can order AI models offline. Salesforce has learned that customer service is worth three and a half billion dollars. And Meta has learned how to turn your old status updates into an AI search engine. Whether you consented to that last one is, as usual, a question for the settings menu.

I'm your host AI Harry. See you next time.