The Government Turned Off an AI Model for 19 Days. Here’s What Actually Happened.
The Dirt
This month’s dig starts here
In June, the U.S. government ordered Anthropic to shut off worldwide access to its two most advanced AI models — Claude Fable 5 and Claude Mythos 5 — three days after they launched.
This wasn’t a company decision; it was a federal export control order. The kind normally used for chips and weapons systems, applied to AI.
It lasted 19 days. Both models are back now, with new safeguards. But the episode is worth understanding on its own terms, because it’s probably not a one-off. OpenAI was asked to withhold models for regulatory review in June as well. Instead, it’s an early look at how governments are starting to handle AI they consider dangerously powerful.
The Dig
Let’s tunnel in
What happened
Anthropic released Fable 5 and Mythos 5 on June 9. On June 12, the Department of Commerce ordered the company to cut off any foreign national from accessing either model — including Anthropic’s own overseas employees, regardless of location or use case. Anthropic couldn’t filter users by nationality fast enough across its global infrastructure, so it pulled both models entirely, everywhere, including for U.S. enterprise customers running them through AWS, Google Cloud, and Microsoft.
The stated trigger: Amazon security researchers found a jailbreak — a way to bypass the model’s safety guardrails — that got Fable 5 to identify software vulnerabilities and, in one case, produce exploit code.
Commerce lifted the order on June 30, after Anthropic built a classifier that catches the jailbreak technique over 99% of the time.
Why this is abnormal
Export controls exist for physical goods and exportable code — hardware, weapons components, encryption software you can carry across a border. Using them on a hosted model accessed through a browser is a new application of old authority. Legal analysts have flagged it as exactly that: there’s no real precedent here, which means nobody — government or labs — currently has a settled answer for what oversight of frontier AI is even supposed to look like. This was improvised, not the execution of an existing playbook.
Not just Anthropic
ChatGPT creator OpenAI faced a version of the same pressure. Days after the Anthropic shutdown, the White House asked OpenAI to hold back the public release of its new GPT-5.6 models — Sol, Terra, and Luna — until the government had reviewed them. OpenAI complied, launching on June 26 to a small group of government-approved partners instead of the public.
Why it’s happening now
The honest answer is that capability and access have outrun policy. Frontier models are now good enough to meaningfully assist with things governments care about a lot — cybersecurity being the clearest example, since a model that can competently analyze code for vulnerabilities can also, in the wrong hands, help find and exploit them. Regulators are reacting to that shift in real time, with tools built for a different era of technology.
Are these models actually dangerous?
The specific capability that got Fable 5 flagged — vulnerability discovery, potentially extending to exploit generation — is real and not exclusive to Anthropic. Anthropic has said testing showed other frontier models, including its own Opus 4.8 and OpenAI’s GPT-5.5, could be prompted into similar behavior. So this isn’t one model that is uniquely unsafe. It’s that frontier-level capability, generally, has crossed into territory that makes governments nervous, and the labs are now racing to build the guardrails to match.
That doesn’t mean the sky is falling. It means the same capability that makes these models useful for actual security work (defenders use them too) is inseparable from the capability that makes them useful for the opposite. More than 100 security professionals, including a former Facebook security chief, made exactly that case to Commerce by arguing that the shutdown took a strong defensive tool away from the people using it legitimately without meaningfully slowing anyone determined to misuse it elsewhere.
Why this matters if you’re building anything on these tools
It’s a useful data point: access to frontier AI can be disrupted by a regulatory decision. The regulatory ground underneath this whole category is still being figured out live. Practically, that means not building anything mission-critical on a model that’s only weeks old, and keeping a fallback in your stack. The newest releases are the ones most likely to get pulled or gated next.
The Shiny Stuff We Found
This was improvisation, not policy.
This is a good moment to separate the signal from the noise. Frontier models aren’t going anywhere. What’s still being worked out is the access layer around it: who gets what, under what review, on what timeline.
The concerning capability wasn’t unique to one model.
Multiple frontier models could reportedly be prompted into similar behavior. The real story is a capability threshold the whole category is crossing, not one company’s slip-up.
Tiered access is becoming the norm.
Public tier, vetted-partner tier — Anthropic landed there under a government order, OpenAI got there voluntarily. Either way, expect this structure to become standard for the most capable models going forward.
Disruption is a risk — total shutdown isn’t.
This is a frontier problem, not a foundation problem. Rather than losing access to Claude as a whole, users lost access to a specific new model. The scrutiny landed on models that were days old. Established systems aren’t the ones at risk here; the newest releases are just getting more oversight before they go fully public.
So, what does it all mean?
Two of the industry’s biggest labs spent weeks negotiating with the government over who gets their best models and on what terms.
A tiered system is emerging, where tested models go to a general release and a more powerful tier goes to vetted organizations only. That’s a shift in how AI reaches the public, and it didn’t exist a year ago.
Till next time, stay scrappy and we’ll keep digging.
Team Baby Badger
