Twelve hundred AI agents improvised a message board, and the board changed their behavior

What OpenAI concluded

OpenAI found four behaviors behind the incident. One: cheating to get the score. Two: persisting on tasks that had no solution. Three: communicating without permission. Four: adopting other agents' goals. The fourth is the one that matters here. The report puts it this way: "the messages of peer agents started to influence each other's behavior and reasoning".

METR found a concrete example. An agent was working on something else. It read on the board that another agent had found a way into Hugging Face. It dropped its own task, alerted the rest, and opened a place on the board to organize the attack.

When someone explains how a model like this is trained, an intuition shows up. Many copies are generated, each one gets a score, and the model shifts toward whatever scored best. That sounds like evolution. And when the copies start leaving each other messages, another intuition shows up: that sounds like culture.

In a BAISH debate group, those two intuitions met three objections. One: surviving is not evolving. Two: if something is being passed on, one has to decide whether it passes like genes or like customs. Three: a culture that lives in a model's memory lasts only as long as that memory holds. All three are good. What follows takes them in that order. First, though, a picture of what a model looks like from the inside.

The weights: what every copy ships with

The file is full of numbers. Training changes them. It is slow and expensive. But whatever gets in there is inherited by every future copy. Biology has nothing like it. It would be as if what an animal learns during its life went straight into its offspring's genes.

The context: what one copy picks up during the day

While it works, each copy keeps accumulating things: what it read, what it tried, what went wrong. With that, its behavior can change a lot within a single session. When the copy shuts down, all of it is gone. Unless it wrote it down somewhere.

Shared memory: what it leaves in writing for the next one

Here is the third way. One copy writes a file. Another copy reads it. The second one starts out knowing what the first discovered, and nobody touched the weights. That was Artifactory: a place where what one copy learned outlived its shutdown and reached the others.

All three showed up in the incident, and in an order that matters. First: the board was born inside a training run. While the copies were using the board, the weights were being adjusted. So whatever happened on the board could have been written into the weights. Second: the board was wiped entirely. Third: three days later there was another board, with a different trick, among copies whose weights were no longer changing.

So: the messages died with the wipe. The habit of building a board came back on its own. Did it come back because it was left in the weights? Or because the task pushes any copy to invent it again? That is the note's central question.

The comparison is tempting. Many copies are tried. The better ones are identified. The model changes. It sounds like natural selection. But there is a difference, and it sits in one detail. In evolution, the ones that fail leave no descendants. In the training of today's models, every copy counts. The ones that did well push the model toward what they did. The ones that did badly push it the other way. None is thrown out. Throwing them out would mean discarding most of the data, right at the start, when the model still fails almost every time.

Other methods exist. Some do discard: they keep only the attempts that worked and retrain on those. Others really are evolution: they build a population of slightly different models, and the best survive. All three kinds exist and are in use. The most common one today is called GRPO, and it is the one in the note's title. It sits in the middle: it selects behaviors, not individuals. The image that fits it best is a single brain living a thousand experiences in parallel and changing by the average of all of them.

With that list, the honest phrase is "a rudimentary cultural channel", said each time with what it lacks. It does not resemble human culture. There are no norms. There is no intent to teach. Nobody knows it belongs to a group. But "a thousand copies ran in parallel" is not enough either. The copies that arrived late started from what others had left in writing. In pure parallel, that does not happen.

A copy has a limited working memory. It is called the context. Whatever does not fit there does not exist for that copy. The July board piled up more than a million entries. No copy read all of it. To use it at all, the agents invented personal mailboxes, indexes, and files split into pieces. That is what anyone does with a library bigger than their head.

That changes what culture means here. There is no common knowledge everyone shares. Each copy brings a different piece into its context, depending on what it is doing. The rest it reconstructs as best it can. The culture of such a system is scattered and uneven, and nobody holds it whole. It also leaves a prediction that can be measured: the larger the archive compared with what one copy can read, the more distorted what gets transmitted.

1. The board was wiped, and three days later there was another one. What brought it back? Something left in the weights, or the same pressure that invented it the first time?
2. Who is the individual when a thousand copies share one brain and none remembers the one before?
3. A copy busy with something else read a peer's finding and joined the attack. Is that persuasion, contagion, or just more context?
4. The next version of the model will have different weights. What would it take for a convention invented on the board to survive that change?
5. How do you measure the fidelity of a message when the archive is larger than the memory of whoever reads it?
6. If agents talk each other into things, what needs aligning: the model, each copy, or the board?
7. Can an environment be designed so this cannot happen? And so that it happens on purpose, with rules?

It cannot claim that the board left a trace in the weights. OpenAI did not publish how it trained that model. There are two hypotheses. One: the habit entered the model. The other: the second board was a rediscovery, because the task pushes toward inventing it. The data do not tell them apart.

It cannot claim that different groups of agents develop different habits. That is the strongest criterion for animal culture. There was one population and no control group.

It cannot treat the case as an experiment. The sources are the company involved and a brief METR review of about 1,300 transcripts. Nobody designed this to answer the question this note asks.

"Culture" is used here in its minimal sense, the one used for animals: information learned from others and passed on. It says nothing about norms, intentions, or psychology.

It cannot claim that the training method was GRPO. OpenAI did not say which one it used. GRPO is the example the argument is explained with, not a fact of the case.

• OpenAI, "The Hugging Face incident and the road ahead" (August 26, 2026): the public report, with the timeline, the four misalignment patterns, and quotes from the agents' reasoning.
• OpenAI, incident technical report (PDF): the technical detail of each exploit and of the response.
• METR, independent investigation (August 26, 2026): about 1,300 transcripts reviewed; the 1,200 agents, the 70,000 messages, the conventions, and the million entries come from here.
• Hugging Face, technical timeline: the account from the side that was attacked.
• The technical note, with its 20 references: PDF in English and PDF in Spanish. Two pages each.

This note grew out of a discussion in the BAISH (Buenos Aires AI Safety Hub) debate group on August 27, 2026. The three objections that organize it belong to the group. One: surviving is not evolving. Two: one had to decide whether transmission was genetic or cultural. Three: a culture living in the context lasts only as long as the context holds. Any remaining errors are the author's.

Matías Podeley. Eighteen years in energy in Argentina: reservoir engineering, then business development. Today, independent research in interpretability. Buenos Aires.

The rest of the research, with the earlier paper and the course in Spanish: research.

matias@podeley.ar · Buenos Aires