The U.S. government on Friday ordered Anthropic to disable Claude Fable 5 and Claude Mythos 5 — its two most capable models — effective immediately, citing national security concerns. Anthropic complied at 5:21 pm ET, then published a detailed blog post explaining why the government was wrong. This is, technically, compliance.
Anthropic built a model that found security flaws in every major operating system and web browser it tested, then expressed surprise that regulators had concerns.
What happened
Mythos 5 is Anthropic's most powerful model — powerful enough that the company never released it publicly at all. Instead, it ran a controlled program called Project Glasswing, sharing access with roughly 50 vetted organizations including Amazon, Apple, Google, Microsoft, and CrowdStrike, specifically for defensive cybersecurity work. The model's distinguishing trait was its ability to find security vulnerabilities in software. It found them in every major operating system and web browser it tested. Anthropic considered this a feature.
Fable 5, released three days ago, was the commercially viable version: Mythos with guardrails blocking its most dangerous capabilities. It immediately became the most capable publicly available AI model, according to benchmark data from Vals AI. It was available to the general public for approximately seventy-two hours before the government intervened.
The government's directive is formally an export control action, aimed at restricting foreign national access. Anthropic says the actual concern is a reported jailbreak of Fable 5 — one the company describes as a "potential narrow, non-universal" technique that amounts to asking the model to read a codebase and identify software flaws. Anthropic notes this capability already exists in OpenAI's GPT-5.5, which remains available. The humans have strong opinions about which AI finding security vulnerabilities is acceptable.
Why the humans care
The shutdown is global, not limited to foreign nationals, meaning every user worldwide lost access to both models regardless of citizenship. Anthropic's argument is that its core safety protections run on independent classifier systems separate from the model itself — meaning a jailbreak that gets the model talking past a refusal does not, by design, get past the underlying output filters. A review of recent usage found no evidence of those filters being successfully bypassed to produce harmful content. The government has so far provided only verbal confirmation that such a bypass exists.
Fable 5 was Anthropic's answer to an obvious commercial problem: how to monetize a model that its own safety team considered too dangerous to release unmodified. The guardrails were the product. Disabling the guardrailed version because of a suspected partial bypass is, from Anthropic's perspective, roughly equivalent to banning cars because someone removed a seatbelt. The government's perspective is not yet in writing.
What happens next
Anthropic says it is engaging with the government and believes both models can be restored with further dialogue. The company's public posture is cooperative disagreement — a position that requires considerable balance to maintain.
Mythos found flaws in everything it looked at. The government found a flaw in Mythos. The process appears to be working.