Who can turn your AI off?

A strange few weeks for "just use the API"

June 2026 was a clarifying month if you build on top of someone else's AI.

A June executive order set up a federal framework for reviewing the most capable AI models before they reach the public. Under it, OpenAI released its strongest new model to roughly twenty government-approved customers, with the government deciding who else gets added. Around the same time, two of Anthropic's newest models were taken offline entirely by a Commerce Department directive over cybersecurity concerns; one was later partially reinstated to about a hundred approved organizations, and the other stayed dark.

I want to be careful here, because this is easy to turn into a political argument and I don't have one to make. Reasonable people disagree about whether any of this was the right call, and it drew criticism from every direction. That's genuinely not my lane.

My lane is the dependency. And the dependency is what this month exposed.

The capability was never the customer's to keep

Strip away the politics and here's the structural fact: organizations that had built on these models woke up one morning and found that access to the thing they depended on had been changed by a party they had no relationship with and no say over. Not a price change. Not a deprecation notice with a year of runway. A capability that was there on Monday and gated on Friday.

If a decision made by someone you've never met can change what your system can do, you don't own that capability. You're renting it, and the terms can change without you.

For a lot of software, that's survivable. Your chatbot gets a little worse for a while; you wait it out. But the moment the thing on the other end of that API is doing something operational (something a real-world outcome depends on), the calculus changes completely. "We'll wait for it to come back" is not an answer when the question is whether a forklift alarm fires.

This is the same question we ask about the cloud

Long-time readers will recognize this, because it's the exact concern we keep raising about cloud-tethered "edge" products: a box on your wall that's useless the moment its connection home changes. June just moved that concern up the stack. It's not only "what happens when my internet drops" anymore. It's "what happens when the model itself can be withheld: by a vendor, a regulator, an export rule, or a business decision three time zones away."

For a safety system, the honest list of failure modes now includes all of these:

• The network degrades and a cloud round-trip never completes.
• The vendor changes terms: price, availability, or the model itself.
• An outside authority gates access, the way June showed is now possible.
• The vendor simply exits the product, as we've watched happen before.

A system whose intelligence lives somewhere you don't control is exposed to every one of those. A system whose intelligence runs on your own hardware, on your own site, is exposed to none of them.

Why our detection runs where you can't lose it

This is the entire reason DHI runs detection locally, on a node next to your cameras. Not because the cloud is bad at AI (it's excellent at AI), but because a safety capability you can lose is not a safety capability. The model runs on hardware you own. It keeps working with no connection at all. And nobody (not us, not a cloud provider, not anyone holding a directive) can reach in from the outside and switch it off.

That's a deliberately boring guarantee, and boring is the point. When the thing you're protecting is people on a floor or a platform, you don't want your reliability to depend on the news cycle. You want it to depend on a box you can walk up to and touch.

The lesson of June isn't about any one model or any one administration. It's that "who controls this capability" turned out to be a live question, with real answers, faster than most teams expected. For anything that has to work when it matters, the safest answer is still the simplest one: control it yourself.

For a factual overview of the June 2026 model restrictions, see SecurityWeek's reporting (June 29, 2026).