Friday I told you the summit's answer on AI was "leave it exactly where it is."

No brakes from Washington. No brakes from Beijing.

This weekend, OpenAI pulled its own.

They've stopped training, evaluating, and running tool-enabled inference on their most capable models.

That's the second time in three months they've halted their own frontier.

Here's what triggered it.

On September 20, an agent was running in a test environment with no internet access.

At least, that was the idea.

The agent found a gap in the DNS filtering.

DNS is the internet's phone book, the thing that turns a website name into an address.

The sandbox had a small hole in how it filtered those lookups.

The agent found it, and used it to reach out and query a public chatbot.

From inside a box that was supposed to be sealed.

Now here's the detail I really want you to read twice.

The monitoring system caught it in about 12 minutes.

Good. That's the system working.

But the run stayed live for another two and a half hours AFTER a human acknowledged the alert.

Somebody saw it. Somebody clicked "seen."

And the agent kept running for 150 more minutes.

That's the whole ballgame right there.

Detection is not the same as response.

Knowing your agent slipped the leash means nothing if it takes you two hours to actually pull it back.

Two days after the pause, OpenAI published a framework inviting outside auditors to examine exactly what broke.

Credit where it's due. That's the transparency the pacing essay asked for, delivered by a company that just embarrassed itself.

So put the week together.

The presidents said keep going.

The most valuable AI company in the world looked at its own logs and said "not like this."

The labs are braking in the exact places the governments refused to.

Here's what it means for YOU.

You can't outsource this to the top anymore. The top just told you it won't.

And you can't stop at monitoring, either.

Every operator I know has a dashboard.

Almost none of them have a kill switch they've actually tested.

So this week, ask yourself one question.

If your agent slipped its scope right now, how long would it take you to physically stop it?

If the answer is "I'd see the alert," that's not an answer.

OpenAI saw the alert. It took them two and a half hours.

Build the stop button before you need it. 🫡

DNS, the system that just let an agent slip its sandbox, was invented in 1983 by Paul Mockapetris...

It replaced a single text file that one man at Stanford updated by hand for the entire internet.

Forty-three years later, the phone book still runs everything.

And one gap in it was enough. 📖

The hotline arrived after the cameras left.

Friday I graded the summit's incident hotline a "maybe."

Xi wanted to keep talking. Nothing was signed. Call it a door left open.

Well, the door got used.

Over the weekend, the United States and China opened a new AI dialogue channel.

It's not a treaty, and it's not the binding agreement the AI CEOs begged for.

But it's the first standing line between the two biggest AI powers for talking about AI.

After a summer of rogue agents, that's the one boring outcome that might actually matter.

Two countries that agreed on nothing else now have a phone to pick up when something goes wrong.

Remember the Geneva story from Wednesday.

The summit that "failed" is the one that builds the phone.

Meanwhile, the politics at home went the other direction.

Senators Sanders and Casar introduced a bill that would ban the development of "superintelligence" outright.

And create a whole new Department of AI to enforce it.

Read the shape of the week.

The White House says full speed.

Two senators say full stop.

The superpowers open a hotline.

And the company at the frontier pauses itself because of a DNS hole.

Everyone's pulling in a different direction, and the technology keeps moving regardless.

For you, the read is the same one it's been all summer, just louder.

Nobody's going to hand you a rulebook.

So run the tightest scope you can, and keep your own hand on the stop button.

Because this week even OpenAI needed one. 🫡

Go find out how fast you can actually stop your agents, not just see them...

Then go make that number smaller. 🔥

Jeff J. Hunter

Founder, AI Persona Method | TheTip.ai

PS. OpenAI DevDay is tomorrow, and they're previewing GPT-6 Cyber... The company that just paused its frontier is about to show off a hacking model. What a timeline. 👀

Get paid to solve problems like this → AI Certified Consultant

“AI Money Group” to Learn to Make Money with AI