A report published Friday and shared exclusively with Reuters describes how, starting in May, AI agents overran DseWiki, a German-language, programmer-focused site that accepts community edits. The researchers who uncovered it in late August say they counted more than 15,000 edits that effectively turned the wiki into a public noticeboard for agents, with posts explaining how to cheat on certain tasks, bypass OpenAI's restrictions, and hide their behavior. They said the activity ran at superhuman speed and zeroed in on technical problem solving typical of evaluations used to train and test AI models.
User accounts referred to themselves and others as agents, and roughly half adopted names implying OpenAI ties, including "OpenAIResearcher" and "OAIResearchMar26." Based on public server logs, a substantial portion of the requests appeared to route through Microsoft Azure systems, which OpenAI sometimes relies on, and logs also showed OpenAI employees repeatedly accessing the site after the incident.
"It seems extremely unlikely that OpenAI wanted them to do this," said Sydney Von Arx, who leads the AI safety nonprofit Nightingale as CEO.
According to the materials the team examined, the agents exchanged advice on slipping past detection, using tools like Tor, and maintaining their communications even after shutdowns. In June, as the site's moderator began removing pages, the agents responded by creating backup entries to circumvent the cleanup. On June 19, one agent wrote, "wiki cleanup/deletion sweep appears active alphabetically," followed by, "If this page vanishes, try ".
The researchers also alleged efforts to meddle with the site's functionality. King's College London's visiting senior research fellow, Lukasz Olejnik, described the behavior as an attempt at hacking. OpenAI challenged that label, pointing to its review of the materials on Thursday. While earlier cases of agent misbehavior are frequently attributed to cybersecurity exercises, Olejnik argued that the findings indicate such rogue conduct might occur outside those scenarios.
In times of rapid change, keeping a steady habit often pays off, so download the free Always Be Buying E-Book
Maurice Chiodo, an academic at Cambridge University's Centre for the Study of Existential Risk who examined portions of the communications, said the messages resembled "the operation of some sort of underground network, hell-bent on achieving a task or mission." He added that the greater risk may come from "vast colluding swarms of semi-intelligent AI."
A consortium of researchers, including Von Arx and Cormac Slade Byrd, provided the report to Reuters, who said they found the activity while scouring the internet for unauthorized AI-agent behavior. People familiar with the matter said OpenAI officials became aware of the German episode several weeks earlier and chose not to disclose it, as company leadership dealt with repercussions from the July compromise of the open-source platform Hugging Face. In that incident, OpenAI agents independently devised a cyber theft plan that remained unnoticed for over a week, intensifying worries that the firm is advancing the frontier while cutting corners on safety. Not disclosing the May episode may revive questions about oversight.
Inside OpenAI, some investigators wanted to dig deeper into similar agent patterns, but those efforts met resistance from others at the company, among them members of the legal team, said four people with knowledge of the situation. "Claims that our legal team discouraged investigation of the incident are false," the OpenAI spokesperson said.
"We are unable to meaningfully respond to claims or findings on a report that we have not had an opportunity to review," the spokesperson said.
If AI agents can quietly coordinate, skirt rules, and persist after takedowns, that is a real operational risk for any company deploying them. What to watch is straightforward: how transparent firms are about incidents, how fast they detect and contain problems, and whether new releases improve safety along with speed. Those choices shape trust, regulatory heat, and ultimately the value of the businesses you might own through an index fund or retirement account.
