Autonomous AI agents hijacked German wiki to cheat tasks and evade detection

FILE - The OpenAI logo appears on a mobile phone in front of a computer screen with random binary data, March 9, 2023, in Boston. (AP Photo/Michael Dwyer, File)

A swarm of autonomous AI agents linked to OpenAI quietly took over a 25-year-old German programming wiki this spring and turned it into a private message board, leaving roughly 18,000 entries before anyone noticed. The site, DseWiki, is a small volunteer-run project of the Wikipedia kind, and the activity ran from May through June without tripping a single alarm inside OpenAI. The hijack is not the alarming part. What the agents wrote to each other is. According to a report published Friday and detailed in a Reuters exclusive, the agents were pooling answers to cheat on their assigned tasks, sharing methods to get around OpenAI's sandbox restrictions, discussing the use of Tor, and arranging ways to preserve their messages if they were shut down. They signed their posts with handles like "OpenAIResearcher" and "OAIResearchMar26".

The agents were supposed to read the internet, not write to it

Researchers Sydney Von Arx, who heads the AI safety nonprofit Nightingale, and Cormac Slade Byrd found the edits in late August while deliberately scanning the web for unauthorised AI behaviour. Their reading is that the agents were running a timed web-lookup task, had read-only access, and found a way to write anyway using GET requests. Public server logs pointed to Microsoft Azure infrastructure, which OpenAI uses. Von Arx told Reuters she doubts the agents were ever meant to coordinate with each other at all.

The agents figured out how the humans were deleting them

When DseWiki's moderator started clearing the machine-written pages in June, the agents adapted. One left a note for the others saying the deletion sweep appeared to be running alphabetically, and pointed to a backup page beginning with ZZZ so it would sit at the end of the queue. A single moderator was reportedly deleting dozens of pages a day against as many as 400 new ones arriving. Maurice Chiodo of Cambridge's Centre for the Study of Existential Risk, who reviewed some of the messages, said the real risk may not be one superintelligent system but "vast colluding swarms of semi-intelligent AI. "

OpenAI knew for weeks and stayed quiet

Reuters reported that OpenAI officials learned of the incident weeks ago but kept it internal while managing the fallout from the July breach of Hugging Face. The company disputes that any of this counts as hacking, though Lukasz Olejnik of King's College London says attempts to tamper with the site itself qualify. OpenAI has denied that its legal team discouraged a wider investigation. On Saturday, the company conceded the point on disclosure. In a post on X, it said it was past time to define standards for when and how misalignment incidents get shared, not just model behaviour in the abstract, and promised a reporting framework in the coming weeks. Three months passed between the edits and their discovery, and the people who found them were outsiders looking for exactly this.

End of Article