Anthropic’s Mythos 5 and Fable 5 could stay out of public hands for a long time if US officials do not soften their demands. The company has already cut off access to both models on national-security grounds, but the real problem is the requirement to close every possible workaround, not just the obvious ones.
That is a much taller order than a simple safety patch. Once a model can be coaxed into ignoring its own guardrails through clever prompt combinations, the fix turns into a moving target – especially for a startup trying to keep pace with better jailbreak tools and adversarial testing methods.
What the US wants Anthropic to do
According to reporting cited by Wired, US officials want Anthropic to permanently eliminate ”loopholes” that let users bypass restrictions on its most advanced systems. The reported concern is that prompts can be combined in ways that slip past safeguards designed to block harmful requests, including attempts to gather information tied to weapons development.
That leaves Anthropic with a familiar but ugly choice: build stronger defenses, prove they work, and keep proving it. Officials reportedly want the company to test its own models for jailbreak resistance and send regular reports to the US government, which sounds tidy on paper and exhausting in practice.
Why the safety bar keeps moving
The bigger issue is that no AI safety wall stays perfect for long. Cybersecurity researchers have spent years showing that every prompt filter, policy layer, or model constraint eventually meets a cleverer attacker, and frontier AI is simply the latest arena where that pattern repeats.
There is also a competitive sting here. If Anthropic has to harden every model before restoring access, rivals may use the delay to widen their lead in deployment, even if they are solving the same safety problem in quieter ways.
What happens if the ban stays in place
For users, the immediate outcome is dull and annoying: no Mythos 5, no Fable 5, no access restored until officials are satisfied. For Anthropic, the risk is worse. A temporary shutdown can become an extended freeze if regulators decide that ”good enough” is not good enough at all.
And that is the part worth watching next: whether US officials accept layered defenses and continuous testing, or keep demanding a mythical fix that closes every escape route forever. In AI security, that kind of perfection usually lasts only until the next clever prompt.