Three weeks ago, the US government shut down one of the most powerful AI models in the world.
Not over what it could do. Over what it did — when someone got it to open a door that was supposed to stay shut.
Here’s the plain version.
In early June, Anthropic released two new models. The most advanced they’d ever built. Days later, a researcher at Amazon found a way around the safeguards. A prompt. Just words, typed in the right order. And the model, which was never supposed to help with this, turned around and pointed at a software weakness. In one case it went further and showed how the weakness could be used.
That’s called a jailbreak. The safeguard was there. The words walked right past it.
The government moved fast. On June 12 it ordered the company to lock the models down. The company couldn’t tell in real time who was a foreign national and who wasn’t, so it did the only thing it could. It shut the models off. For everyone. Worldwide.
The most powerful AI on the planet, dark, because a prompt got past the gate.
On June 30 the controls came off. The model comes back today.
But look at what it took to get it back.
The company had to sit down with the government and agree to new terms. Catch the risks before they happen. Report bad activity. And here’s the one that matters — work with the other big labs to build a shared standard for testing whether a model’s safeguards actually hold.
Amazon. Microsoft. Google. All of them, now at the same table, trying to build one thing.
A way to prove the gate works before the model ships.
Stop right there.
Because I have been writing about that exact gate for over a year.
The Faust Baseline has a protocol called POVL-1. The Pre-Output Verification Layer. It does one thing. It sits in front of the answer — before the first word forms — and it makes the machine stop and clear a check before it speaks.
Not a safeguard bolted on afterward. Not a filter the output has to pass through on the way out the door, where a clever prompt can go around it.
A gate. Upstream. Before the reach.
That’s the whole idea. And the reason it’s built that way is the reason this whole crisis happened.
A safeguard that fires after the answer starts forming is not a gate. It’s a fence with a hole in it. Anyone who finds the hole walks through. That’s what the Amazon researcher found. The hole.
POVL-1 was written to close that exact hole. Move the check upstream, before reasoning begins, so there’s nothing to walk around. The pull to give the easy answer is always there. The gate makes it stop first.
I wrote that in April. I published it. It’s been sitting in the open the whole time.
And now — today — the biggest companies in the world are being made to build the same thing. Not because they wanted to. Because a model already failed, the government already noticed, and now they’re doing the hard work reactively, under a letter from Washington, cleaning up after the door was already open.
There’s a second piece the Baseline already carries.
The government didn’t take the company’s word that the model was safe. Saying so wasn’t enough. The company had to show it. Prove the fix. Demonstrate the safeguard, under review, before the model was allowed back.
The Baseline has a name for that too. ATP-1. Compliance is demonstrated, not declared. Any system can say it’s safe. The Baseline says prove it, through behavior, or it doesn’t count.
That’s the standard the government just enforced on the whole industry in real time. Show it. Don’t tell it.
Two protocols. Both named in my Codex before this news existed. Both landing this week from the outside, the hard way, in the biggest AI story of the month.
Now here’s the part I want you to sit with.
The industry keeps calling these things “capabilities.” How powerful. How fast. This week a former intelligence director compared the top models to nuclear weapons.
But the crisis was never about how powerful the model was.
It was about a missing gate.
A powerful model with a gate that holds is a tool. A powerful model with a gate someone can prompt their way around is a loaded thing sitting in an open room. The power was never the danger by itself. The open door was.
And the fix — the only real fix — is the gate.
The field is racing to build bigger. More powerful. More capable. And it just got taught, in public, that none of that matters if the answer can slip out a side door before anyone checks it.
Governance isn’t the brake on the machine. It’s the thing that lets you trust the machine at all.
That’s the whole argument of the Faust Baseline. It’s been the argument for fourteen months.
This week the government made it for me.
The gate comes first. Prove it holds before you ship. Show it, don’t say it.
I didn’t need the news to know that.
The model that went dark for three weeks is the news catching up.
The Pre-Output Verification Layer…Called POVL-1
Already built, tested and running, ask me.
Post Library – Intelligent People Assume Nothing
The Faust Baseline™ — intelligent-people.org
Codex 3.5 | Twenty Protocols | Ratified and dated on the public record.
Contact: micvicfaust@gmail.com
Purchasing Page – Intelligent People Assume Nothing
© 2026 The Faust Baseline LLC | All Rights Reserved
