Anthropic defends models, says flaw also affects GPT-5.5
After the US government ordered access to Fable 5 and Mythos 5 cut off for all foreign nationals, Anthropic says the flaw is narrow and exists in the publicly available GPT-5.5 too.
- The US government ordered Anthropic to block access to its Fable 5 and Mythos 5 models for all foreign nationals, including foreigners working at the company inside the US.
- Anthropic says the flaw is narrow and that the same result can be drawn from OpenAI's GPT-5.5, which is not under the same controls.
- The trigger was a claim that another team "jailbroke" Mythos; Anthropic says it received only verbal evidence.
- Fable 5 launched on June 9 and the model is used by several hundred million people.
What happens when the government tells a company to switch off its own product in the middle of the day? That is exactly what hit Anthropic this week, and the company does not agree with the reason.
The US administration ordered Anthropic to disable access to its most capable models, Fable 5 and Mythos 5, for all foreign nationals. The measure makes no distinction by location. It affects foreigners employed by Anthropic itself on US soil, as well as users outside the country. The stated justification is national security.
Anthropic responded publicly that it believes this is a misunderstanding and that it is working to restore access. What makes the story interesting is the company’s argument: the flaw that caused the alarm is not exclusive to its model.
Why does Anthropic say the ban makes no sense?
According to the company, the trigger was another team’s claim that it had “jailbroken” Mythos, meaning it bypassed the model’s safeguards. Anthropic says it received only verbal evidence of a “narrow, non-universal” flaw from the government. In its view, that is not enough reason to pull a model already used by several hundred million people.
This is where the main argument comes in. Anthropic says the same technique can be applied to other publicly available models, specifically OpenAI’s GPT-5.5, which faces no similar export controls. In other words, if the capability is dangerous, it is not locked behind a single product. Competing models, the company says, show a similar ability to find minor bugs in code.
Mythos has been described as a model able to find security flaws in every major operating system and browser it was tested against. That sounds powerful, but it is also a capability that, by Anthropic’s argument, no single system on the market holds alone.
Some context is worth adding. Fable 5 launched on June 9, priced at $10 per million input tokens and $50 per million output tokens. Before the order, it is reported, Amazon chief Andy Jassy warned the White House, and administration adviser David Sacks claimed Anthropic failed to act on a warning about a possible jailbreak.
Is this about security or pressure on one firm?
AI policy expert Dean Ball summed up a doubt shared by many in the industry. He said he could not tell whether this was a legal attack aimed specifically at Anthropic or extreme national-security hawkery, but that the move was cartoonish either way.
That is the heart of the dispute. If the technology is genuinely dangerous, the logic of export controls calls for applying them to everyone who holds it, not just one player. If it applies to only one, the question stops being technical and becomes political.
For users this is not an abstract story. Models that half the world already uses for work suddenly become a zone of regulatory uncertainty, where access depends on citizenship rather than a contract or a subscription. Anthropic says for now that it is negotiating to restore access, but the precedent remains.
Which is more dangerous to you: a tool that finds a hole in a system, or a rule that targets a single company while the same capability sits freely somewhere else?
