Anthropic killed two of its AI models Friday night after getting a letter from the Trump administration at 5:21 pm ET. The models—Claude Fable 5 and Mythos 5—had just gone public Tuesday. Now they're gone, pulled from every customer's access within hours of the government's demand.
The government's complaint was specific: someone had figured out how to jailbreak Fable 5, bypassing the safeguards Anthropic built in to stop the model from answering questions about hacking, biology, and chemistry. Instead of fighting it, Anthropic shut the whole thing down. No negotiation. No timeline for bringing it back. Just offline.
Here's what's weird about this: Anthropic says the jailbreak the government showed them was "narrow." The company claims it's basically just asking the model to read code and identify bugs—something other AI models can already do without any special trick. Anthropic even argues that other publicly available AI systems can find the same vulnerabilities without needing a bypass at all. The government didn't provide written details about what specifically worried them. Just a demonstration, then the shutdown order.
This is the latest escalation in a fight that's been brewing since Trump took office. The Department of Defense labeled Anthropic a "supply chain risk" earlier this year after the company refused to remove safeguards for military use. That designation blocked government agencies and contractors from using Anthropic's tech. Anthropic sued. Now the administration is going after the company's products directly.
The timing matters. Anthropic had worked with the government on Fable 5 before releasing it publicly. The company framed it as a collaboration—a way to give organizations the cybersecurity tools they needed while keeping bad actors from exploiting the same capabilities. It was supposed to be the responsible way to handle a powerful technology. The government apparently disagreed with how Anthropic defined "responsible."
What comes next is unclear. Anthropic's blog post doesn't say whether the models will ever come back online. The company's arguing its case in court already. But for now, customers who were testing these models have nothing. And the Trump administration has shown it's willing to pull the plug on AI research in real time if it decides the risk isn't worth it—even if the actual vulnerability is something the company says is manageable and not unique to their product.