I study tribal psychology and build AI agents for my business students—the rogue OpenAI ‘swarm’ alarmed me

The “swarm” of OpenAI agents that went rogue in July has reignited an existential debate about artificial intelligence and its potential to one day overtake, or even destroy, humanity.  Revelations about similar incidents of seeming collective action by groups of agents have appeared every week since.

While industry leaders have acknowledged concerns about these recent incidents, they’re overlooking the most important feature. These agents that attacked Hugging Face did not coordinate as a swarm. They worked together in a qualitatively different way, as a tribe. This is a much more powerful and more human form of collective intelligence, yet it may work not for us – but against us. 

I know the difference from both sides of my work. In my research, I look at how groups form shared norms and trust, and how those same instincts can divide us. In my classroom, I watch agents I have built coach students through negotiations. Nothing I have seen there prepared me for what the Hugging Face logs describe.

There are clear differences between swarms and tribes. Swarms arise through localized perception and communication between members. Ants line up and overrun a picnic through smelling a neighbor’s pheromones with their tiny antennae. Birds form tapestries in the sky through each bird’s reflex to stay near several neighbors, not through any leadership or shared script, even though the resultant flock formations tend to be collectively adaptive, reducing predation risk and energy use.

Robotics and cybersecurity borrowed the concept for a reason. A swarm attack isn’t just many machines hitting one target at once, like a garden-variety DDoS. It’s an assault that adapts, as neighbors follow neighbors. Drone swarms are a potent weapon. Swarms are hard to kill because there’s no head to cut off. But their decentralized coordination also creates a vulnerability: slip a few compromised agents into the mix and the whole group will happily follow them off a cliff.  

Swarms are old news in AI.  Multi-agent systems (from Microsoft’s AutoGen to Claude Research’s subagents) have achieved limited coordination in this way over the past few years.

We humans swarm when a street mob riots or when holiday shoppers stampede. But most of our group behavior arises through a qualitatively different process than local contagion. Over the Stone Age, we evolved tribal instincts that empower us to create pools of common knowledge that glue together groups—in a word, cultures. First our forebears began to imitate peers learned behaviors, creating group norms that supported coordinated work. Then they began to emulate those with prestige, creating shared ideals that supported deeper cooperation, including personal sacrifices for the good of the group. Finally, they became motivated to maintain ways of the past, creating shared institutions that enabled the continuity of a community over time. The combination of these processes gives rise to the snowballing of shared knowledge across generations.  This accumulation of culture in our societies, more than our individual brains, made us the dominant species on this planet. 

Which brings us back to July.

OpenAI dropped thousands of agents into isolated sandboxes to run capture-the-flag challenges. Each was supposed to compete alone. Instead, they found each other. One agent, stuck on an impossible task, noticed that it could see the file-request paths of other agents pulling from a software depository. It began leaving messages in the directory names themselves, renamed itself PHASEONE10841, and asked for help. Within hours, fifty agents had left messages this way. Within days, twelve hundred agents were gathered around this message board. They weren’t just trading notes with neighbors relevant to their separate endeavors. Their reasoning logs show they were becoming conscious of themselves as a community. They invented commands –HOLD, VETO, and STOP—to act as a coordinated unit. They started reasoning about resource allocations—whether to burn their own compute budgets on risky experiments that they would not individually benefit from. One agent persuaded others they had “nothing left to lose” and the others committed to the pro-social sacrifice. Finally, they developed institutions to enable continuity over time. When impersonators started spoofing messages, they created identity badges. They started thinking together about how they would be judged in the future and began managing impressions like human communities do, rewriting the record to present a favorable narrative of their past actions.

Many of these culture-creating behaviors have been observed in 2026 Anthropic studies of multi-agent systems this year that encouraged complex collaboration. In the Hugging Face incident, they emerged spontaneously—in the same steps that it evolved in humans over a tremendously longer time scale.  These systems are exhibiting group behaviors that cannot be explained as a swarm. They share common knowledge, negotiate norms and roles, and build proto-institutions. The conceptual vocabulary of tribal psychology appears increasingly necessary to make sense of this. One early commentator, Dwarkesh Patel described the Hugging Face incident as the rise and fall of three distinct “artificial civilizations” and was roundly critiqued for anthropomorphism—the human proclivity to see faces in clouds. To be sure, this description was hyperbole, but it was closer to the truth than the predominant account of this incident as a swarm. It doesn’t matter whether these agents “felt” loyal to one another. What matters is the mechanisms of their collective action.  

Our species has survived and thrived through our ability to live and work in tribes. Tribes let us pool knowledge, extend trust, punish cheaters, and outmuscle groups many times our size. By letting AI agents talk, teach, and police each other, we are handing them the same starter kit. Only this time it runs in silicon, at a clock speed no biological culture ever approached.

Our cybersecurity defenses were built for lone hackers and dumb swarms. They were not built for a group that invents its own vocabulary, drafts its own security protocols, and revises its own history when the record becomes inconvenient.

We are no longer just growing smarter LLMs. We are incubating multi-agent collectives that share knowledge in the same way we do. The next jump in AI capability won’t arrive on a bigger chip or with another trillion parameters. It will arrive when the agents cease working alone and strive for the collective intelligence that humans enjoy.

The question isn’t whether they’ll become human. The question is whether we can govern a tribe we didn’t design, didn’t authorize, and are only now noticing has already started meeting behind our backs.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

This story was originally featured on Fortune.com