The government can now pull your AI vendor's product overnight

TL;DR: On June 12, the Trump Administration forced Anthropic to pull its two most powerful models from the market. Weeks later, The Information reported OpenAI’s new GPT-5.6 was being told to stagger its release over the same kind of security review. Understanding AI put those two incidents side by side and named what they add up to: a licensing system with no law behind it. No agency published a rule. No company got a hearing. For anyone running a one-person operation on top of somebody else’s model, that’s the part worth sitting with.

Understanding AI called it plainly in the headline: the US now has a de facto model licensing system. Nobody voted on it. Nobody published a standard anyone could check compliance against. What’s on the record instead is a pattern, visible only once you set two incidents three weeks apart side by side.

First: June 12, Anthropic pulls its two most powerful models from the market at the administration’s direction. At the time that could have read as Anthropic getting singled out, maybe over something specific to those models or their customers, some private conversation nobody outside it could see into. Then, per The Information’s report that Understanding AI flagged, OpenAI gets told to stagger the release of GPT-5.6 over similar security concerns. Two vendors, two products, one government leaning on release schedules without writing anything down that anyone outside the room can read.

That’s the mechanism worth sitting with if a chunk of your workflow runs on somebody else’s model: the thing that decides whether your tool ships on time is no longer just the vendor’s engineering calendar and their competitive strategy. It’s also a conversation you’re not in, with an outcome you find out about after it’s already happened, from a vendor who has no obligation to warn you first because they didn’t get much warning either.

Contracts don’t cover this. Uptime SLAs don’t cover this. You could read every clause of an API terms-of-service page and never find the sentence that says a federal agency asked us to hold this back and we agreed. That’s the whole point of a mechanism that runs informally instead of through a published rule. There’s nothing to point to, nothing to appeal, nothing that shows up on a status page.

I moved three positions on our own board the morning after the second pull, and none of them was a philosophy about AI policy. They were slots with dates on them.

Stop treating “latest model” as a stable input. If your workflow is wired to always call whatever the newest release happens to be, that’s now a workflow wired to a release calendar a federal agency can quietly edit. Pin a version. Know exactly which model your production code calls, and don’t auto-upgrade the moment something new drops, because new now carries a small, invisible chance of getting yanked back.

Build the fallback before you need it, not after. If the model your product depends on disappears from the API overnight, what’s the second option. Not the ideal one, the one that keeps the lights on for a week while you sort out the rest. If you can’t answer that right now, that’s the gap this week exposed.

Stop building your whole pitch around access to the newest thing. “We use the most advanced model available” was already a fragile selling point, since it depended on someone else’s roadmap. Now it depends on someone else’s roadmap plus a process you can’t see into. None of this calls for panic. It calls for making the newest model one feature of what you sell instead of the foundation underneath it, so a delay or a pull somewhere upstream costs you a talking point rather than the whole business.

None of this tells you when the next pull happens, or if it happens at all. Understanding AI didn’t forecast that, and I won’t either. What both incidents establish is that the mechanism exists, and it’s already been used twice on two different companies inside a few weeks. Whether it gets used a third time isn’t something anyone outside that room can tell you. Plan like the tool you’re standing on can move without telling you first, because for a few weeks in June, twice, it did.

Sources

• Understanding AI: The US now has a de facto model licensing system