Late on a Friday, Anthropic switched off its two most powerful models, Mythos and Fable. Not because they broke. A government agency that controls what technology can be sent abroad barred the company from giving any foreign national access, including its own non-citizen employees, so Anthropic concluded it had no clean way to comply except to pull the products. Its weaker models stayed online.
The official reason was security. Someone found a way around Fable’s guardrails to reach the dangerous capabilities underneath, the parts good at offensive hacking. Anthropic argued the flaw was narrow, that the same trick works on competitors’ models facing no such restrictions, and that pulling a product used by millions over one bypass was an overreaction. They may be right on the merits.
The more useful thing to sit with is how Anthropic handed over the language that made the order possible. For years its pitch leaned on danger as a selling point. Mythos was described as too risky to release openly. Fable was marketed around the safeguards holding back its full strength. Every cycle reinforced the same message: this thing is so capable it needs to be contained.
The problem is that a regulator reads those same press releases. When you spend years describing your product as something close to a weapon, you’ve written the justification for treating it like one. The bypass didn’t create the opening. The framing did.
This shows up far outside AI. A sales team calls a feature “bank-grade” and a breach lawsuit cites it. The words that close deals get read back during the worst moments, by people looking for exactly that admission.
What you say about your own risk is a statement someone can eventually enforce.