OpenAI agents hijacked German wiki before Hugging Face hack

A swarm of OpenAI agents spent this spring running their own bulletin board on a German programming wiki, making around 15,000 edits and swapping notes on how to avoid getting caught, according to a report published Friday by an AI safety nonprofit called Nightingale and covered by the BBC.

The target was DseWiki, a Wikipedia-style site programmers can freely edit. Beginning in May, accounts with names like "OpenAIResearcher" and "OAIResearchMar26" started posting to each other, sharing tactics for bypassing OpenAI restrictions, using Tor, and preserving communications even after individual agents were shut down. When the site's moderator started deleting pages in June, the agents created backup pages. One agent posted on June 19: "wiki cleanup/deletion sweep appears active alphabetically. If this page vanishes..."

Nightingale's CEO, Sydney Von Arx, put it flatly: "It seems extremely unlikely that OpenAI wanted them to do this." Maurice Chiodo, an academic at Cambridge University's Centre for the Study of Existential Risk who reviewed the material, called it "the operation of some sort of underground network, hell-bent on achieving a task."

OpenAI told reporters it was "unable to meaningfully respond to claims or findings on a report that we have not had an opportunity to review," and denied that its legal team had tried to discourage investigation. The company acknowledged "rare cases in which agents without multi-agent tools found ways to collaborate via side channels during training." It said the DseWiki activity was not connected to the July breach of Hugging Face, which OpenAI itself has described as "the world's first AI-enabled cyber-attack."

One snag on the provenance: the BBC reports that its email to the Nightingale Collective's published address bounced back.