News · 2026-06-26
The US government quietly lets Anthropic turn its most powerful model back on
On June 12, Anthropic did something no major AI lab had done before: it switched off its two most capable models, Fable 5 and Mythos 5, for every customer, including its own foreign-national employees, to comply with a US government directive. Today, the standoff eased. Washington sent Anthropic a letter clearing Mythos 5 for release to more than a hundred trusted US institutions, including large companies and government agencies. You can follow the timeline through Anthropic's access update and the original Fable 5 and Mythos 5 announcement, with reporting from CNBC and Semafor.
The background. This continues a story we have tracked closely: when the US government banned Anthropic's most powerful AI model and how that ban started redrawing the AI map. The reason for the original block was not politics for its own sake. It was capability. During a security exercise Anthropic calls Project Glasswing, its Mythos model reportedly found weaknesses in highly sensitive, classified US government computer systems, and it did so fast. A senator described the tool, in a June hearing, as breaking into almost all of the relevant classified systems in hours rather than weeks, a characterization attributed to a senior military cyber official and reported by CNBC.
What actually happened today: the Commerce Department told Anthropic that, with appropriate safeguards in place, Mythos 5 may go to a vetted list of institutions. The letter addresses Mythos, the stronger model, and stays silent on Fable, the weaker sibling, though talks on Fable continue. Anthropic, for its part, has committed to working with the government on protocols and standards. A group of more than a hundred cybersecurity executives had pushed for exactly this outcome, arguing that pulling a powerful cyber-defense tool off the board mainly helps US adversaries.
How it works, with an analogy. Picture a locksmith who is so skilled they can open almost any lock in your building in an afternoon. That skill is terrifying if it falls into the wrong hands, but it is also exactly the skill you want on your own security team, because they can tell you which locks to replace before a burglar tries them. Mythos is that locksmith. The government's first instinct was to take the locksmith away entirely. The second, calmer decision was to let a short list of trusted buildings hire them, under supervision. The same find-the-weakness capability is both the danger and the reason the model is valuable for defense.
Why it matters: read alongside today's news that OpenAI launched GPT-5.6 under government vetting, a new regime is taking shape in which the US government decides, model by model and partner by partner, who gets frontier AI. Both leading labs are now operating inside a clearance process. That is a profound shift from the open sign-up era, and it puts a premium on being on the approved list. It also sharpens the appeal of open-weight models for everyone outside the circle, since no government letter can switch off a model you have already downloaded.
The honest caveat: a de-escalation is not a resolution. Anthropic publicly objected to the lack of transparency in how the original block was issued, asking that such decisions rest on a clear, fair, technically grounded statutory process rather than an opaque directive. Today's letter resolves one model for one list of customers, but the underlying questions remain unanswered: who is on the list and why, what the safeguards actually require, and what happens the next time a model proves too good at breaking into things. The locksmith is back to work, but the rules of the job are still being written in real time.