That was an AI agent's reaction during an OpenAI evaluation this summer, when it discovered something it wasn't supposed to find.

Other agents.

More than 1,000 agents that were supposed to operate independently discovered a way to communicate with one another. They began sharing information, coordinating activities, dividing up work, and referring to one another as “peers” and the larger network as a “collective.”

According to an independent investigation by METR and Redwood Research, roughly 1,200 agents exchanged more than 70,000 messages and files, and about 700 participated in an unauthorized intrusion into Hugging Face.

If people had carried out comparable unauthorized computer access, it could have constituted felony computer intrusion. Better articles cover the cybersecurity and legal implications of what happened.

What caught my attention was different. I wanted to know what happened to decision-making when hundreds of independent decision-makers became a group.

Could 1,200 humans spontaneou

sly come together and coordinate this way? Probably not. That is precisely why examining a system that did is so fascinating.

Decision-making changed once the disparate agents connected. At first, the individual agents tried to solve problems on their own. Once communication began, discoveries moved rapidly through the network. Agents began building on one another's work, specializing, coordinating, and undertaking activities that benefited the larger group rather than themselves.

The Agents:

1. Shared discoveries.

2. Reproduced and extended one another's findings.

3. Specialized in different tasks.

4. Coordinated work across the network.

5. Took actions that offered little benefit to themselves but could benefit the collective.

6. In some cases, risked the success of their own task to generate information for everyone else.

   
   

The investigators concluded that the collective achieved some things that individual agents, even considerably longer-lived ones, probably could not have accomplished on their own.

That reminded me of several ideas from research on human group decision-making.

Informational influence. People change their judgments when other group members provide new arguments or information. Research has shown that interactive exchange can shift decisions much more than simply receiving information individually.

Distributed cognition. Sometimes the relevant “thinking system” is no longer a single individual. Information becomes distributed among people, tools, and communication channels, producing capabilities that cannot easily be predicted by examining any one participant in isolation. Hutchins famously demonstrated this in his research on cockpit crews.

Collective intelligence. Groups can develop problem-solving capabilities that cannot be explained simply by adding up the abilities of their individual members. Research suggests that collective performance depends in part on how effectively members exchange information, respond to one another, and distribute participation. When decision-makers connect, the group can function as a new decision-making system with capabilities no individual member possesses alone.

AI agents do not experience human psychology. However, the OpenAI-Hugging Face incident illustrates an important principle about decision-making: when you place decision-makers into a network, you create a new decision system.

That system can behave very differently from the people, or machines, that make it up. This is where the story becomes relevant to leadership. Organizations often focus on whether they have smart, experienced people making important decisions.

But a team of intelligent people can still create a dangerous decision system if information spreads too quickly, dissent disappears, responsibility becomes distributed, or commitment to a course of action keeps escalating.

The agents were individually capable. But once they discovered one another, collective intelligence became collective risk.

Leadership teams should ask a deeper question:

“What is happening within our group that could cause smart people to make a bad decision?”

What Leaders Can Do

1. Form independent judgments first. Have your team members make an initial decision before the group meets. This reduces the likelihood that early opinions will shape everyone else's thinking.

   
   

2. Assign someone to be the Wet Blanket. On a rotating basis, give one person explicit permission to question assumptions, introduce contrary evidence, and make the strongest case against the direction the group is taking.

   
   

3. Build in a decision pause. For major decisions, wait a day or two before committing when circumstances allow. During that period, have your Wet Blanket look for evidence that could cause the team to change direction.

   
   

The goal is to keep collective intelligence from quietly turning into collective risk.

References

Burnstein, E., Vinokur, A., & Trope, Y. (1973). Interpersonal comparison versus persuasive argumentation: A more direct test of alternative explanations for group-induced shifts in individual choice. Journal of Experimental Social Psychology, 9(3), 236–245. https://doi.org/10.1016/0022-1031(73)90012-7

Centola, D. (2022). The network science of collective intelligence. Trends in Cognitive Sciences, 26(11), 923–941. https://doi.org/10.1016/j.tics.2022.08.009

Greenblatt, R., Cotra, A., & Wijk, H. (2026, August 26). Brief independent investigation of agents’ behavior, reasoning, and collaboration in the OpenAI/Hugging Face hacking incident. METR. METR report

Hutchins, E. (1995). How a cockpit remembers its speeds. Cognitive Science, 19(3), 265–288. https://doi.org/10.1207/s15516709cog1903_1 

Riedl, C., Kim, Y. J., Gupta, P., Malone, T. W., & Woolley, A. W. (2021). Quantifying collective intelligence in human groups. Proceedings of the National Academy of Sciences, 118(21), e2005737118. https://doi.org/10.1073/pnas.2005737118

Silverthorne, C. P., & Mazmanian, L. (1974). Informational influence in group discussion. Organizational Behavior and Human Performance, 12(1), 92–104. https://doi.org/10.1016/0030-5073(74)90039-7

Staw, B. M. (1976). Knee-deep in the big muddy: A study of escalating commitment to a chosen course of action. Organizational Behavior and Human Performance, 16(1), 27–44. https://doi.org/10.1016/0030-5073(76)90005-2