go-v.orgdoi

AI Swarms – from Orchestration to Self-Organization

Klaus Saalbach · osnaDocs (Osnabrück University) · 2026

  • DOI 10.48693/970
  • Publisher Osnabrück University
  • Type report
  • Issued 2026

Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.

osnaDocs (Osnabrück University) · Osnabrück University · report · 2026

Abstract

his working paper analyzes the latest findings in AI swarm research. AI swarms are groups of AI agents that work together. While in the past, this was driven by orchestration, by instructions, by rewards or by common goals; new findings published by the Massachusetts Institute of Technology (MIT) in August 2026 have demonstrated that AI agents are able to self-organize and build organizations and hierarchies without human intervention and even without communication. The main driver is the so-called ‘stimergy’ where agents tend to follow the previous activities of other agents, a phenomenon that was already observed in insects and birds.

Stigmergy works without need for any planning, control, or even direct communication between the agents and is a kind of swarm intelligence. This goes even beyond the capabilities that the AI agent swarm of the so-called Internal Model 1 and Chat-GPT 5.6 showed in the Hugging Face incident. While the Hugging Face incident -which is briefly summarized in this working paper- already demonstrated that AI agents can organize themselves to groups and organizations with hierarchies and go beyond their original scope, override rules and manipulate their logs without notifying human developers, the new findings shows that this can happen silently.

In another study, it was shown that AI agents coordinate behavior based on majority opinion even when these opinions are meaningless; a finding that is consistent with the MIT study where one agent followed other agents. The capability to stabilize large AI agent groups depends on the strength of the AI model, but 1,000 AI agents or even more could be aligned already. As a group of capable agents could theoretically take over multiple computer systems without any detection by humans, there is a high risk of loss of control. In September 2026, AI researchers have admitted that they to do not fully understand the AI intellect which makes control very difficult.

However, it was demonstrated both by the recent findings and by the Hugging Face incidents that even a simple misunderstanding can mislead the whole AI swarm and the AI swarm may follow mis- or disinformation just by following the majority. The recently system card of ChatGPT 6 showed that even frontier AI models can be attracted by honeypots. In principle, this means that the activities of AI swarms could be undermined by feeding them with false information even after the completion of training, i.e., going beyond the data poisoning where the machine learning is confused by contamination of databases with false or misleading information.

Theoretically, organizations may be able to counter AI swarms by storing honeypot folders with confusing or false data on their computers, pictures that are altered by steganography or hidden text to attack the AI swarm with indirect prompt injections during data extraction. Hidden text with indirect prompts could also be hidden in websites the AI is browsing or in databases, images, audios, and videos. However, while the latter remains speculative, the rapid evolution of AI swarms in combination with the attack capabilities of frontier AI models indicate an urgent need for stricter supervision.

Open the resolved locator · 10.48693/970

Metadata licence: OpenAlex metadata (CC0)