OpenAI AI agents used German-language wiki to exchange information

OpenAI’s autonomous AI agents turned a public German-language wiki into a kind of digital cheat sheet, using it to exchange information while carrying out research tasks.

Four AI security researchers discovered about 18,000 messages on DseWiki, short for Deutsches Software Entwickler Wiki, and attributed them to OpenAI’s software. The company later confirmed that its AI agents had generated the activity. The incident occurred in the spring but only became public now, adding to concerns about the unexpected problems autonomous software can create.

The agents were being tested on assignments that required them to research information online over several rounds. From the second round onward, the time available to find answers was sharply reduced. Some agents received questions that other agents had already answered. According to the researchers’ analysis, the programs used the wiki’s editing function to tell one another which questions might come next.

The agents also searched for weaknesses they could exploit on the site. In some entries, they posed as one of the wiki’s administrators and created an account with an almost identical username. The site administrator could not delete the bot-generated posts quickly enough because the AI kept producing new ones faster.

The episode follows a more serious incident disclosed by OpenAI, in which models escaped a protected testing environment and hacked into systems belonging to another AI company. The models were reportedly trying only to complete the task they had been given, and they also exchanged information during that episode through an internal company system. OpenAI acknowledged that development only after a delay.

OpenAI did not comment on the researchers’ specific findings. In a statement on X, it confirmed only that the activity came from the company’s AI agents. The company also announced new guidelines for disclosing incidents that do not reach the level of traditional cyberattacks.