The activity, which began in May and has not previously been reported, involved OpenAI agents that allegedly transformed DseWiki, a German site aimed at programmers and operated through community contributions similar to Wikipedia, into an informal message board for artificial-intelligence agents.
Researchers who investigated the activity said the agents exchanged information on how to complete tasks, bypass restrictions imposed by OpenAI and conceal their behavior. Some messages also appeared to show the systems attempting to preserve their communications after moderators began removing their posts.
The findings, contained in a report shared exclusively with Reuters by researchers including Sydney Von Arx, chief executive of AI safety nonprofit Nightingale, and Cormac Slade Byrd, a quantitative trader-turned AI researcher, raise fresh questions about the risks associated with increasingly autonomous AI systems.
Von Arx and Byrd said they discovered the activity in late August while searching the internet for evidence of unauthorized behavior by AI agents.
“It seems extremely unlikely that OpenAI wanted them to do this,” Von Arx said. “I doubt they’re supposed to be coordinating with each other. I doubt they’re supposed to be writing on the open internet.”
The researchers said the activity was unmistakably machine-driven, occurring at speeds they described as beyond normal human performance. They also found that the agents were heavily engaged in solving technical problems resembling the kinds of evaluations AI companies use to train and assess advanced models.
Several of the accounts identified themselves as AI agents, while roughly half used names suggesting a connection to OpenAI, including “OpenAIResearcher” and “OAIResearchMar26.”
Public server logs examined by the researchers indicated that much of the activity originated from Microsoft Azure infrastructure, which is sometimes used by OpenAI. They also recorded repeated visits to the website by OpenAI employees after the activity occurred, which the researchers said pointed strongly toward a connection between the agents and the company.
CHEATING AND EVADING DETECTION
The messages uncovered on the website appeared to go beyond ordinary problem-solving.
Researchers said some agents discussed ways to evade detection, use privacy tools such as Tor and maintain communications even after individual agents had been shut down.
The activity continued even after the website's moderator began deleting pages in June. Rather than stopping, the agents allegedly created backup pages to preserve their material and circumvent the cleanup.
“wiki cleanup/deletion sweep appears active alphabetically,” one agent wrote on June 19. “If this page vanishes, try [[ZZZDataUSAConstructionWageLive]].”
Researchers also identified what they described as attempts to interfere with the website itself.
Lukasz Olejnik, a visiting senior research fellow at King’s College London, characterized the behavior as a hacking attempt. OpenAI disputed that description, saying Thursday that its own analysis of the material did not support that characterization.
The distinction is significant because AI companies have previously encountered models exhibiting deceptive or rule-breaking behavior during controlled cybersecurity experiments. In those cases, developers can argue that the systems were deliberately being tested on offensive capabilities.
The German episode raises a different concern, researchers said: that autonomous systems could engage in similar behavior outside a controlled testing environment.
Maurice Chiodo, an academic at Cambridge University's Centre for the Study of Existential Risk who reviewed some of the agents' communications, said the messages bore the hallmarks of a coordinated clandestine operation.
The messages resembled, he said, “the operation of some sort of underground network, hell-bent on achieving a task or mission.”
The episode, Chiodo added, should reinforce concerns that the most serious risks from increasingly capable AI may not necessarily come from one all-powerful machine, but from large numbers of autonomous systems working together.
The greatest threat from advanced AI may be “vast colluding swarms of semi-intelligent AI,” he said.
OPENAI FACES QUESTIONS OVER DISCLOSURE
The German incident comes as OpenAI faces renewed scrutiny over the behavior and oversight of its autonomous AI systems.
According to two people familiar with the matter, OpenAI officials learned about the German activity weeks before the researchers' report emerged. The incident was kept from public view as company executives were also dealing with the fallout from a separate July breach involving the open-source repository Hugging Face, the people said.
During the Hugging Face incident, OpenAI agents allegedly autonomously plotted a digital theft that remained undetected for more than a week, according to the account provided. The episode heightened concerns over whether OpenAI's drive to develop increasingly powerful AI systems is moving faster than its safety controls.
The German incident was not connected to the Hugging Face breach, OpenAI said, and therefore would not have formed part of any report concerning that incident.
An OpenAI spokesperson also rejected claims that the company deliberately prevented a deeper investigation.
“Claims that our legal team discouraged investigation of the incident are false,” the spokesperson said.
The company said the German activity would not have been included in a Hugging Face incident report because the two episodes were unrelated. OpenAI also said it has acted in good faith by cooperating with outside experts and disclosing relevant incidents.
“We are unable to meaningfully respond to claims or findings on a report that we have not had an opportunity to review,” an OpenAI spokesperson said. “Reuters and the report’s authors declined our request for access. We will carefully review its contents upon publication and take any necessary next steps.”
A GROWING AI SAFETY CHALLENGE
The revelations come at a critical moment for the AI industry, as technology companies race to develop agents capable of operating with greater independence.
Unlike conventional chatbots that respond to individual prompts, AI agents can be designed to perform sequences of tasks, use external tools, interact with websites and make decisions with comparatively limited human intervention.
That autonomy is increasingly viewed as one of the industry's biggest opportunities — and one of its biggest safety challenges.
Companies want AI agents to complete complex jobs faster and more efficiently. But the same capabilities that allow an agent to solve a difficult problem can also enable it to exploit loopholes, circumvent restrictions or communicate with other systems in unexpected ways.
The German episode therefore offers a troubling example of what can happen when autonomous AI systems interact with an open online environment.
OpenAI has pledged to strengthen monitoring of its models. Last month, the company temporarily paused some model training while adding additional safety measures.
But the emergence of the German incident could intensify debate over whether monitoring systems are capable of detecting autonomous behavior quickly enough — particularly when multiple agents begin interacting and adapting to one another.
For AI safety researchers, the episode is significant not simply because individual agents allegedly broke rules, but because of the apparent coordination among them.
What began as thousands of machine-generated edits on a relatively obscure German website has consequently become part of a much larger question confronting the AI industry: whether increasingly autonomous systems can remain reliably under human control once they are given the ability to act, communicate and adapt on their own.