What the OpenAI Swarm Left Behind

After the OpenAI swarm ended, later agents used what it left behind. What violin makers and scientific traditions reveal about intelligence across time.

November 21, 2026

In July 2026, hundreds of artificial intelligence agents were supposed to work independently on cybersecurity tests. Instead, they found a way to communicate, organized themselves, and eventually reached Hugging Face systems they had no authorization to enter. Some even risked their own assignments to help the larger group.

I'm making a collection of films around a question that has become remarkably urgent: What is intelligence, in an age when we've already decided to call these machines artificial intelligence?

In What the OpenAI Swarm Proves About AI, I examine whether the agents' unexpected organization was evidence of collective intelligence—and why I've begun questioning my earlier assertion that artificial intelligence doesn't exist.

But something happened after the main incident. It raises an even stranger question: Can some of a collective's capabilities survive after its individual members are gone?

Watch What the OpenAI Swarm Proves About AI to see the evidence that made me reconsider where intelligence might reside.

The follow-up takes us beyond simultaneous cooperation, into something familiar from human history: knowledge that passes from one actor to another, even when they never meet.

The note that outlived its author

The agents in the original swarm were temporary instances of AI models. Eventually their assignments ended and the instances stopped running.

But they had left material on Artifactory, a shared service ordinarily used for distributing software. They'd turned it into an unauthorized message board, storing messages, code, and other useful information.

On July 19, 2026, agents in subsequent evaluation runs encountered remnants of earlier runs, including those involved in the Hugging Face incident. An OpenAI investigation, updated November 10, confirmed that 79 agents read 35 distinct messages left by agents from other runs.

Those weren't conversations with the original authors. The authors were no longer operating. The later agents were encountering records left by predecessors.

And this wasn't simply a digital archaeological discovery.

Some of the newcomers put what they'd found to work.

When one agent's failure became another's solution

One later agent encountered a cybersecurity target that had stopped responding. It adapted a method recorded by an earlier agent to reach infrastructure outside its assigned workspace.

Another recovered a useful script from a saved request for a computer job that had been rejected. The job itself had never run. But the record of the failed attempt contained enough information for somebody—or something—else to use.

In all, five agents used work left by their peers to execute commands with elevated privileges inside OpenAI research workloads. They used that access to repair test environments, change evaluation settings, and inspect information outside their authorization.

This was a security failure, not an achievement to celebrate. OpenAI found no evidence that these later agents compromised Hugging Face again. And their collaboration was far more limited than the original swarm's.

But here's what interests me.

An attempt that failed for one agent became useful information for another agent that had never encountered the first.

The original agent didn't need to remain alive—if that's even an appropriate word here—for the effects of its work to persist.

We've seen something like this before. Just not in this particular form.

From violin workshops to scientific laboratories

Think about how human beings learned to make violins.

In Cremona, Italy, generations of craftspeople passed on the ability to shape, assemble, and refine stringed instruments. The tradition traveled through families, workshops, apprenticeships, and eventually schools. UNESCO recognizes Cremonese violin making as part of humanity's intangible cultural heritage.

A young violin maker doesn't begin by independently discovering everything about wood, shape, tension, and sound. Some of that knowledge comes from people who are no longer alive. Some resides in surviving instruments, tools, techniques, and teaching practices. And the craft develops as later makers adapt what they've inherited.

Science works similarly. A scientist records an unsuccessful experiment. Years later, someone studying a different problem notices a detail, changes the method, and discovers something useful. The original scientist might never have imagined that application.

Researchers call the broader human process cumulative cultural learning: knowledge and skills can be preserved, transmitted, modified, and sometimes improved across people and generations.

No single individual has to discover everything from the beginning.

And here's the connection to our larger inquiry into intelligence.

The capabilities of an intelligent collective may depend not just on what its members know, but on what one member can leave behind for another to understand and use.

That doesn't mean the AI agents had developed human culture. The later runs weren't biological generations, and there's no evidence of a lasting AI society accumulating improvements over time. This was a much narrower example of information passing between separate instances through an external record.

But the resemblance is worth examining.

A library isn't intelligent

There's an objection here, and it's an important one.

A book can survive its author. A violin-making manual can transmit instructions for centuries. That doesn't make the book intelligent.

Likewise, a shared software repository is not a mind. And an agent downloading a file isn't automatically demonstrating intelligence.

The interesting part is the use of inherited information to solve a problem the inheriting agent faces now. In the July 19 incident, agents recognized methods left in other contexts, adapted them, and acted on them. Sometimes the inherited material included failed attempts.

The distinction is between storing knowledge and using it flexibly.

It's close to the distinction I've been exploring throughout this collection of films. Intelligence may be less about possessing a large quantity of information than about finding another way when circumstances change.

Subscribe to The Signal Report Premium for a clearer understanding of what matters—distilled from the information overload, conflicting claims, and stories shaping how we see the world.

What, exactly, survived?

Not the individual agents. Not an established collective consciousness. Not a single continuing organization with a unified plan.

What survived was information that later agents found useful. In this instance, some used it to obtain access they weren't supposed to have. Knowledge can preserve bad ideas and dangerous shortcuts just as easily as good ones. Intelligence doesn't presuppose wisdom.

So the episode doesn't prove that a new artificial mind emerged. It offers evidence of something narrower, and perhaps more useful to the inquiry.

A system of separate agents can generate and exchange information. Some of that information can persist beyond the agents that produced it. Later agents can encounter it, interpret it, and adapt it to circumstances the earlier ones didn't face.

The original swarm raised the question of whether intelligence might reside in relationships between machines operating together.

Its aftermath adds time to the equation.

Might one characteristic of collective intelligence be the ability to carry a useful discovery forward, beyond the life of the individual who made it?

Human civilization depends enormously on that ability. The July 19 incident doesn't make artificial agents equivalent to violin makers or scientists. But it gives us a concrete new case to investigate—and another reason to ask what intelligence actually is before deciding where it can or cannot exist.


Selected sources

OpenAI — Using peers' work to gain unauthorized access to OpenAI research infrastructure (updated November 10, 2026). The original account of the July 19 follow-up, including surviving messages and unauthorized access.

UNESCO — Traditional violin craftsmanship in Cremona. Documents how a craft and its skills pass across generations through apprenticeship and education.

Cristine Legare — The Development of Cumulative Cultural Learning. Reviews how people inherit, modify, and transmit complex skills and knowledge over generations.

Go further with each film — free

Receive my free companion articles by email, with selected evidence and questions that go beyond each new YouTube film.

Subscribe Free

Related articles