Anthropic didn’t want this fight. The company had been trying to smooth things over with Washington after a messy fallout earlier this year. But on Friday, the US government handed down an order that left no room for negotiation: disable your best AI models. Now. For everyone.
The directive came wrapped in the language of national security. The Pentagon’s chief information officer said some things matter more than “revenue cycles” and “pre-IPO valuation.” Translation? Anthropic’s upcoming stock offering won’t save it from government pressure. Washington sees a threat, and it’s pulling the plug on two of the company’s most advanced models—Fable 5 and Mythos 5—for all foreign nationals.
What sparked this? A narrow vulnerability. A potential way to trick the safeguards built into Fable 5, forcing it to help identify software flaws it wasn’t supposed to touch. Anthropic got only verbal evidence of this “jailbreak,” no formal documentation, no detailed security briefing. Just a government demand to shut it down.
A Company Caught Between Two Worlds
Anthropic’s problem started when it refused to help the US military. The company didn’t want its AI used for domestic surveillance or fully autonomous weapons. That decision cost them. Washington responded by putting Anthropic on a supply chain blacklist, penalties set to kick in later this year.
Now comes the export control hammer. The company must immediately disable access to hundreds of millions of deployed AI models because of one potential jailbreak that the government acknowledges is “narrow” and “non-universal.”
Anthropic’s pushback is worth reading. The company said the government is acting without proper safeguards, without transparency, without the kind of regulation it had actually been calling for. Just days before this order, Anthropic publicly urged Washington for stronger oversight of AI. They wanted rules. They wanted clarity. They didn’t want this—a sudden, sweeping ban based on a possibility rather than a proven threat.
When Safeguards Become Targets
Experts worry that Mythos-class models in the wrong hands could accelerate sophisticated cyberattacks. Banks. Interconnected systems. Decades-old technology running critical infrastructure. A powerful AI that knows how to find weaknesses could cause real damage. The government’s concern isn’t abstract—it’s practical.
But Anthropic says rival AI providers show similar abilities to find minor bugs in code. So why just them? This marks a major shift in how America is controlling AI—not just the chips and tools that power it, but access to the software itself.
The question hanging over this isn’t about whether the threat is real. It’s whether one potential vulnerability in one model justifies pulling the plug on hundreds of millions of users. Anthropic disagrees. The Pentagon doesn’t care what Anthropic thinks. The order stands. The models are coming down. And the race between AI companies and government regulators just entered a new, messier phase.
