Software companies have spent decades deciding for themselves when a product is ready. That assumption did not survive June. When Anthropic released its two most capable models, Mythos 5 and Fable 5, the US government stepped in within days, suspended access, and put the systems through technical reviews across several federal agencies before letting them back online. By the time the models returned, in staged approvals stretching over more than two weeks, the story had stopped being about one company's launch. It had become the clearest sign yet that the release of a frontier AI system is turning into a government-reviewed process.
The trigger, according to reporting collected by IBTimes, was mundane in isolation: a reported jailbreak vulnerability, flagged when Amazon's chief executive raised concerns with the Treasury. What followed was not. Anthropic sent engineers to Washington to argue the flaw was already fixed, and federal reviewers, including the National Security Agency, were unconvinced and demanded more. The company eventually shipped protections it says block the technique most of the time, committed to deeper collaboration with agencies, and agreed to help build a shared framework for evaluating powerful models before they ship at all.
What makes the episode a precedent rather than a one-off is how quickly it rippled outward. OpenAI, the reporting notes, released its own GPT-5.6 into a narrow preview for a handful of organisations rather than a broad rollout, saying the limited launch was coordinated with the government while a review framework is written. That framework traces back to a June executive order directing agencies to build a formal process for vetting advanced systems with significant cybersecurity capability before wider deployment. Two of the largest labs have now, in effect, cleared their most important products with Washington first.
Read alongside a separate thread of coverage, the shape of the new bargain comes into focus. The Times of India reported that even as the Commerce Department lifted its export controls on Anthropic's models, the two sides had not discussed the government taking an equity stake, the arrangement OpenAI has floated to ease political scrutiny. In other words, the price of getting back to market was not ownership. It was compliance: new safety classifiers, ongoing agency access, and a standing right for the government to reimpose restrictions if the company falls short of its commitments.
There is a genuine case for all this. Frontier models are edging into territory, particularly in cybersecurity, where a single capable release could do real damage, and a pre-launch look is a reasonable insurance policy. But the tensions are just as real. A review process with no published rulebook hands regulators enormous discretion over which products reach the public and when. Anthropic's own documentation shows Fable 5 now carrying safety classifiers that can decline requests outright, while its restricted sibling Mythos 5 does not, a two-tier world in which the most powerful version is reserved for a vetted few.
The open questions are the ones the framework has not yet answered. Who decides when a model is safe, against what standard, and with what appeal? A smaller lab without engineers to fly to Washington may find the bar effectively higher than a well-resourced incumbent can clear with ease. For now, one thing is settled that was not a month ago. Releasing a frontier model is no longer solely a product decision. It has become, quietly, a question of national security policy.