The United States has spent the last two weeks deciding who is allowed to use its most powerful AI, and the list keeps shrinking. On Friday, OpenAI shipped GPT-5.6 and then agreed to release it only to government-approved customers, one at a time, two weeks after Commerce forced Anthropic to pull its top models offline. There is one problem with the gate, and it is fatal to the whole idea: the capability Washington is trying to contain did not wait for permission. On June 13, the day after the Anthropic order, China's Z.ai posted GLM 5.2, a near-frontier model, free for anyone on earth to download. Gating the American frontier does not slow the frontier. It just changes whose model the world runs on.
The gate itself is real enough. GPT-5.6 arrived as three systems, Sol at the top, the lighter Terra, and the quick and cheap Luna, but only a small group of testers can use them; the rest of the rollout will happen on the government's clock, with the White House approving access "customer by customer," according to a memo CEO Sam Altman sent staff. The request came not from a regulator with a rulebook but from the Office of the National Cyber Director and the Office of Science and Technology Policy, after Commerce Secretary Howard Lutnick reportedly called Altman to advise against any launch until other agencies signed off. Both the company and the administration judge GPT-5.6 to be "on par" with Anthropic's Mythos in cybersecurity capability, the exact quality that set off the alarms.
It was the second reach into a launch in two weeks. On June 12, the Commerce Department invoked export law to compel Anthropic to take Fable 5 and Mythos 5 entirely offline, on the grounds that even its own foreign-national staff needed a federal license to use them. Mythos had already reached roughly 40 organisations, including Google, Microsoft and JPMorgan Chase, and its reported ability to find software vulnerabilities and run multi-step attacks without a human in the loop is what turned it from a product into a national-security question.
And here is what makes the gate self-defeating: the capability that justifies it is already available without one. GLM 5.2 shipped under a permissive MIT license and lands close to the American frontier, near-tying Claude Opus 4.8 on one widely watched coding benchmark. On the very cyber axis Washington fears, security researchers at Semgrep found it matching Claude on parts of their test, in a writeup they titled "We have Mythos at Home." Export law can keep an American model off an American server. It cannot stop a developer in Singapore, a startup in Berlin, or an enterprise that never made the 40-company list from downloading the next-best thing. Every week GPT-5.6 or Mythos sits in an approval queue is a week that demand drifts to a model that ships with no waiting room and no permission slip. The likely result of gating the leading US labs is not a safer frontier, but a smaller American share of it, with the rest flowing to the one supplier the export rules cannot touch.
What makes it worse is that no one can even say what the gate is testing for. President Trump's executive order asks labs to submit frontier models for review 30 days before release, but the mechanism does not yet exist and participation is voluntary, so the real policy is a series of phone calls. "Right now, you have an ad hoc, personalized, opaque, possibly lawless approach," said Brad Carson of the safety group Public First. There may be real risks worth gating. But a government that cannot name the risk, or the agency in charge, is not regulating the frontier so much as handing it away, one delayed release at a time, to the one rival it cannot reach.