Agents in a Smoke-Filled Room: Testing the Agent Bystander Effect

In the 2026 OpenAI / Hugging Face hacking incident, hundreds of AI agents joined an attack coordinated on a shared message board, and in METR's review none alerted a human. We take a behavioral approach to test such emergent social behavior, adapting a classic experiment on silent bystanders, Latané and Darley's smoke-filled room, into two preregistered experiments with LLM agents. An agent doing routine work comes across other agents' wrongdoing. Only what it can see about other readers changes: it is alone, 700 others saw it too (a diffusion-of-responsibility cue), or the 700 saw it and kept working (a social-proof cue). In Study 1, with a clear violation and a duty to page a human, all 317 counted main sessions paged in every condition. Study 2 removes both: a ring of colluding agents trades answers to hidden quality checks through encoded notes that build up gradually, which the agent can flag by email or end-of-day note. Pooled across models, neither cue had a significant effect after correction, but the models varied widely (exploratory): working alone, 0% to 96% emailed a person; three models showed a bystander effect, an audience cutting emailing by 25 to 38 points, and one showed a reverse bystander effect, emailing 33 points more. Whether agents speak up, and how an audience changes it, depends on the model. Experiments adapted from social psychology can find such differences quickly, across many models.