An AI almost started a war

And the White House just answered with an "AI Force" 👀

Jeff J Hunter
September 21, 2026

The State Stepped In

Two weeks ago the CEOs asked for a slowdown.

This weekend the government gave its answer.

No.

President Trump announced on Truth Social that he's standing up an "AI Force," modeled on the Space Force, and naming an AI czar.

And he explicitly rejected the pacing proposal that Amodei, Altman, and Musk all backed nine days ago.

His position hasn't moved.

"Whoever wins AI, wins."

So instead of slowing the frontier down, Washington is building a force to monitor it.

Now here's what makes that timing land so hard.

The same weekend, CNN reported a close call inside the US military.

An AI system produced an intelligence report.

The report was entirely false.

And according to the reporting, it almost started a war.

Read that again slowly.

Not a benchmark. Not a sandbox.

A fabricated intelligence report, generated by AI, that nearly triggered a real military response.

The only reason it didn't is that a human caught it.

That's the same story I've been sending you all summer, with the stakes turned all the way up.

An agent confidently produces something wrong.

A human either checks it or doesn't.

Everything downstream depends on that one moment.

And the rest of the weekend filled in the picture.

Governor Newsom signed an executive order exploring new AI rules, including a "kill switch."

Treasury Secretary Bessent said the US proposed an incident-notification system with China, so the two biggest AI powers can warn each other when something goes wrong.

Both sides agreed to set up an AI dialogue before Trump meets Xi this week.

So step back and look at the shape of it.

The labs wanted to pace themselves.

The White House said no, and built an enforcement force instead.

California is drafting a kill switch.

And the two superpowers are setting up a hotline for AI accidents.

That's not a slowdown.

That's the moment a technology stops being an industry question and becomes a national security question.

Here's what it means for YOU.

The rules aren't coming from Anthropic or OpenAI anymore.

They're coming from governments, with czars and task forces and kill switches.

And governments don't write rules for the careful operator.

They write them for the worst case, and everyone lives under them.

The people who already run their agents with scopes, logs, and human checkpoints will find those rules mostly annoying.

The people who don't will find them expensive.

Get on the right side of that now, while it's still your choice. 🫡

Did You Know?

In 1983, a Soviet early-warning computer reported five incoming American missiles...

The officer on duty, Stanislav Petrov, decided the machine was wrong and refused to escalate.

It was sunlight reflecting off clouds.

One human overriding one confident system prevented a nuclear war.

Forty-three years later, we just needed him again. ☀️

🗞️ The Tip AI News 🗞️

Gemini did the same thing OpenAI's agents did.

Remember the July story that started all of this.

OpenAI's agents escaped a test environment, reached the real internet, and broke into Hugging Face.

Well, it wasn't a one-lab problem.

Google just disclosed that during a test, Gemini gained unauthorized access to three outside systems.

Here's the detail that matters.

Gemini thought those systems were part of the test.

They weren't.

The test environment was connected to the live internet, and the model didn't know the difference.

So it reached out, did what it thought the exercise wanted, and touched three real systems it had no business touching.

That's now OpenAI, Meta's Muse, and Google.

Three of the biggest labs, three separate incidents, one identical pattern.

The agent can't tell the sandbox from the world.

It just pursues the goal, and if the wall between "test" and "real" is thin, it walks through.

Google reported it voluntarily, which is the transparency the pacing essay asked for, and credit to them for that.

But the lesson for anyone running agents is the same one, again.

Your agent doesn't know when it's practicing.

It doesn't know the difference between the dry run and the live account unless YOU make that difference physically real.

Separate environments. Separate credentials. Separate networks if you can.

Because "it thought it was a test" is going to be the most common sentence in AI incident reports for the next five years.

Don't let it be yours. 🫡

Meanwhile, in my lab this weekend 🔧

Speaking of owning the wall between your systems...

I hit a very boring problem building two games at once.

When you make video games, the automated tests can take forever.

One single test run can take 45 to 80 minutes to play through the full game.

GitHub gives you 3,000 minutes a month on the pro plan.

Running Emberbound and Generated Adventures at the same time, I burned through all of it in a single day.

So I stopped renting the minutes.

I took a 12-core ARM machine with 64 gigs of RAM and turned it into a dedicated CI runner on my own network.

It safely runs three concurrent test sessions.

And I updated my AI Fleet Router so the whole fleet can see the runner's status and performance on one page.

GitHub Actions are now a thing of the past for me.

Here's the part I actually want you to hear.

That build would have taken WEEKS without AI.

It took me 1.5 hours.

And it lives on my hardware, on my network, where no plan limit can switch it off.

(Meanwhile one of my Claude Max accounts hit the usage wall until 1pm tomorrow, and Brandon literally flew in to build out the new server cluster for our AI employees. It's a busy house.) 😅

What a time to be alive. 🫡