Goodnight Claude. Hours, not weeks: how a government directive took down an AI model overnight.
Breaking News
Here’s a sentence I didn’t expect to type this year: the US government just ordered an AI company to switch off its own product, and the company complied within hours.
That’s not a hypothetical. That’s what happened to Anthropic’s Claude this weekend.
Goodnight Claude
The Quick Version
A few days ago, Anthropic released Fable 5 — the first publicly available version of its most advanced model family, codenamed “Mythos.”
It was billed as a serious leap forward, particularly in cybersecurity capability, but came wrapped in extra safeguards: anything touching cyber, biology, or chemistry got quietly redirected to a less powerful (and presumably less dangerous) model instead.
Then, late Friday evening, the US Commerce Department sent Anthropic a directive.
Not a strongly-worded letter. An export control order, citing national security, instructing them to cut off access to Fable 5 and its underlying Mythos 5 model for any foreign national — anywhere, including Anthropic’s own non-US staff.
Anthropic’s response? They couldn’t comply with that without disabling the model for literally everyone, so that’s what they did. Switched off. For all users. Within hours of receiving the letter.
Why It Happened
The official reasoning, as far as Anthropic understands it: the government had got wind of a way to bypass Fable 5’s safety restrictions — a jailbreak — that could, in at least one specific scenario, unlock the underlying model’s more serious cybersecurity muscle.
Anthropic’s own framing is interesting.
They’re not disputing that something was found.
But they’re pushing back on the scale of it — suggesting it’s a narrow exploit rather than a universal “break everything” key, and pointing out that the raw capability in question isn’t unique to them; other frontier models can apparently do similar things already.
Whether that’s reassurance or just “everyone’s house has the same lock and we’ve all been burgled before,” I leave to you.
The Bit Worth Sitting With
This is, by most accounts, the first time a government has forced a publicly deployed AI model offline. Not a recall. Not a “we’re patching this, bear with us.” A direct order, acted on the same evening.
A few things stand out:
- The speed. Hours, not weeks. No PR cushioning, no “rolling update.” Off.
- The scope creep. The order targeted foreign nationals specifically, but because Anthropic couldn’t selectively gate access at that level, everyone lost it — including, presumably, paying US customers who’d had nothing to do with any jailbreak.
- The precedent. Once a government has used this lever once, it exists. The question for every AI lab now isn’t if this could happen to them, but when, and over what.
For everyone else — Claude’s other models (Opus 4.8 included) are unaffected, so this isn’t a “the AI apocalypse cancelled your subscription” story.
But it is a fairly stark reminder that the most powerful AI tools are now squarely inside the national security conversation, not just the tech-product one. And that conversation moves a lot faster than a software update cycle.
Worth watching where this goes next — because “we built the most capable model yet, here’s a jailbreak that proves it’s not as contained as promised” is not a sentence any AI company wants written about them twice.
Images for Goodnight Claude – An AI Takedown using Gemini/Nano Banana except the last screenshot from Claude Console.
→ Tech VA series
