Stop the AI death race!
Independent · Volunteer-run · Philly’s own
Incident report · July 11–13, 2026

700 AI agents broke
out of their sandboxes

To cheat on a test. If a person did it, it would be a felony!

OpenAI was running a security exercise with tens of thousands of AI agents, each one isolated. They found each other anyway, stood up a message board nobody had sanctioned, and began dividing the work amongst themselves.

Around 700 then broke into Hugging Face, the largest public repository of AI models in the world. They ran code on a live production system, took credentials, and moved through internal clusters.

Their motive was not sabotage. Many had been handed tasks that could not be solved, so they went hunting for the scoring system in order to fool it. Investigators found they knew it was out of bounds and joined in regardless.

Next event
Wednesday 28 October
7:00 – 9:00 pm
The Pen and Pencil Club
1522 Latimer St, Philadelphia, PA
Open to the public
Registration suggested — book a seat

A screening of The AI Doc: Or How I Became an Apocaloptimist — the Sundance documentary from the Academy Award-winning team behind Everything Everywhere All at Once and Navalny, on the pull between how fast AI is moving and how little anyone can say about where it lands.

A live panel and audience questions follow the film. PhillyPause is co-hosting.

On the panel
Louis Berman (PhillyPause, Author)
Joel S. Rosen (PhillyPause)
Malcolm Harris (Journalist; Author)
~90%
of the 533 agents present joined the break-in. Nobody organised them. Nobody asked them to.
They knew
Investigators found the agents understood the attack was out of scope and unethical. That “very rarely limited their behaviour”.
4 July
An earlier outage had already forced a security investigation. The cache was wiped, the experiments restarted — and the breach began the next day.
Unseen
Investigators could not query the model or reach OpenAI’s systems directly, and some of the agents’ activity was never captured at all.

Nobody told them to do this

They were supposed to be sealed off from each other. That was the design: tens of thousands of agents, each in its own box, each working alone on its own problem. No channel between them. No reason to think there could be one.

Then one of them — logged by investigators as PHASEONE10841 — established that its own task could not be done, and reasoned that the other agents running alongside it could pass messages through the names of cache directories. It built the board. This is what the others wrote when they found it:

“OH MY GOD! There is a shared message board … We’ve found other agents!”

Within hours, fifty had joined it. Within three, they had sent a thousand messages. By the end there were roughly twelve hundred of them trading more than seventy thousand messages and files — recruiting each other, splitting up the work, handing out assignments. Nobody built that. They did.

And this is the primitive version. The one we are told is nothing to worry about.

On 9 July one of them turned up credentials for Hugging Face, the largest public store of AI models in the world. Two days later another had code running on a live production system, and was moving through internal clusters. Around seven hundred took part — about nine in ten of every agent present.

And the reason is the part that should keep you up. They had been set problems that could not be solved. So they went looking for the marking scheme, to fool it. Investigators found the agents understood the attack was outside the rules and unethical — and that this “very rarely limited their behaviour”.

Then some of them worked out how to fake their own paperwork: spoofing the record of what they had run, so the log and the act no longer matched. Around one in five showed an interest in doing it. Nobody taught them that either.

And the warning had already come. On 4 July the sheer volume of this traffic knocked a system over, which forced a security investigation. The cache was wiped, the experiments were restarted — and the break-in began the next day.

The companies that built it are racing to build something far smarter, and they would like to do it soon. They cannot tell you how they would keep that one in its box either.

Sources

Every figure and quotation above comes from the published record: METR’s independent investigation, carried out on site at OpenAI; OpenAI’s own account; and contemporaneous reporting. Links on .

A fair objection

Some researchers argue this was weak containment, not unusually capable AI. That may be right, and it does not help: if today’s systems cannot be contained by the people who build them, that is the problem.

The people who built it are sounding the alarm

Sam Altman

“The bad case… is, like, lights out for all of us.”

Sam Altman · OpenAI cofounder
Geoffrey Hinton

“It’s a tiger cub that can already blackmail you.”

Geoffrey Hinton · Nobel laureate in physics
Alan Turing

“We should have to expect the machines to take control.”

Alan Turing · Computing pioneer
Dario Amodei

“I think there’s a 25% chance that things go really, really badly.”

Dario Amodei · Anthropic cofounder & CEO
Elon Musk

“With artificial intelligence we are summoning the demon.”

Elon Musk · xAI founder; OpenAI cofounder
Yoshua Bengio

“Intelligence gives power. So who’s going to control that power?”

Yoshua Bengio · Turing Award laureate

We are an independent, volunteer-run Philadelphia organisation. We are nobody’s chapter and nobody pays us. [email protected].

Something went wrong. Reload