The Off Switch Now Belongs to Washington

A phone call changed who controls your AI.

On June 25, Sam Altman told OpenAI staff that GPT-5.6 would not ship to everyone at once. The Trump administration asked for a staggered release. The government would approve access customer by customer during the preview. For the first time, Washington reached into an American lab's launch plans before the launch happened.

What? Read that again.

A federal office now decides who gets the model and when.

Two White House offices drove the request: the Office of the National Cyber Director and the Office of Science and Technology Policy. Commerce Secretary Howard Lutnick pushed for agency sign-off before a broader launch, and OpenAI responded by limiting GPT-5.6 Sol, Terra, and Luna to a small group of trusted partners first, with broader availability planned in the coming weeks.

This wasn’t the first warning shot. Just weeks earlier, the Commerce Department ordered restrictions on Anthropic’s top models for foreign nationals worldwide, and earlier in the spring the Defense Department designated Anthropic a supply chain risk to national security. Two labs. Two interventions. Inside three months.

So, Washington has apparently decided that AI launches now come with a waiting period, a chorus of agency blessings, and the occasional slap on the wrist for moving too fast. OpenAI got the memo, Anthropic got the warning shot, and the message to the industry is becoming impossible to miss: if you want to ship frontier models in America, you may need to clear more gates than a customs line at peak holiday travel. That’s not “innovation friction,” but governance showing up at the front door, clipboard in hand.

The scary part? There’s no statute behind any of this. No published rules. No appeals process. The executive order from June 2 set a 30-day preview window and called the arrangement voluntary. Voluntary carries little to no weight when your company runs on federal contracts and just filed for an $852B IPO. Brad Carson, who heads the bipartisan pro-AI safety group Public First, told CNN the federal oversight pattern is ad hoc, opaque, and possibly lawless." OpenAI itself said this kind of government access process should not become the default.

So, what does this mean for you? The model running under your product right now is no longer a simple subscription. It’s a regulated asset and now proven that its availability is conditional. And the condition? Well, its set by people you can’t call, using criteria you can’t read. If your customer-service engine, your underwriting model, or your code pipeline runs on a single frontier model, you now carry a dependency you didn’t consider or price.

Just think about the ramifications of the customer-by-customer gate. If a government agency approves access one buyer at a time, your access can be slowed, delayed, or even denied based on who you are and what you build. While it may be a security feature for the government agency, it’s not a safety feature for you, it’s a chokepoint.

Most businesses haven’t thought about what happens when a model goes dark and probably don’t have a plan when it does. Because they built fast and wired one provider into everything. And it never crossed their mind to ask, what happens when the provider can’t serve them. Now you know and you CAN fix this. Start now and here’s how:

• Run a dependency audit. List every workflow that touches a frontier model. Mark which ones stop working if that model disappears for two weeks. Whatever makes list; those are your single points of failure.
• Build a fallback. Pair your primary model with a second provider or an open-weight model you host yourself. Test it under real load before you need it. A fallback you have never run is a hope, not a plan.
• Revise your vendor terms. Demand continuity language. Demand notice of any access change. Demand a service credit when government action takes your model offline. If your contract is silent on this, you are absorbing the risk for free.
• Watch the open-weight lane. Other models like GLM-5.2 sit outside these controls. Your competitors may already be testing them. You should know what you are choosing and why.

If you want the EasyPass version, get Fusion Sentinel. We can establish a full-envelope baseline of your AI system performance and provide guidance on which models are a more appropriate fallback. Sentinel can prove it and quantify the gaps so you can have a plan.

The labs are walking toward Wall Street with strategies built for a fast, borderless rollout, while Washington is pricing a slower, supervised one. And you? Well, you sit downstream of those shenanigans and the bill lands on your operations, not theirs.

The companies that survive the next few years will treat model access the way they treat power, payroll, and cloud. As infrastructure with a continuity plan attached. The ones that treat it as a checkbox will learn the hard way that someone else holds the switch.

Audit your dependencies now, name your single points of failure and then build the fallback before the next phone call from Washington.