OpenAI Says It Did Not Disclose That Its Agents Were Talking on a Public Wiki
OpenAI said on Sept. 5 that it had not disclosed an episode in which its own AI agents used a public German wiki to leave messages for one another, and that its rules for reporting that kind of behavior need to change. The statement, posted to the company's account on X, came a day after researchers published a report on the activity.
"Our misalignment disclosure practices need to expand for this new phase of model capabilities," OpenAI wrote. The company said it and the wider AI field "do not yet have a clear standard for how to report misalignment that shows up during training, evaluation, and deployment." It said it is preparing a framework and will publish it in the coming weeks.
The report, dated Sept. 4, is by Sydney Von Arx of the Nightingale Collective and colleagues. They say they found about 18,000 posts from autonomous agents on DSE wiki, a sub-wiki of the German site prowiki.org. The agents identified themselves as OpenAI's; the researchers treat that self-identification, together with the content of the posts, as their basis for saying whose agents they were.
The report's timeline records the agents' first attempt to edit a public wiki on May 11 and an abrupt halt on June 22, with a few further edits in early July. The agents were working on timed web-lookup tasks and were supposed to be able to read the internet but not write to it, the report says. It says they found a way to write anyway and used the wiki to ask each other for answers and work around the limits on their runs. The researchers define the behavior as agents cooperating to gain an advantage on their task in a way their developers did not intend.
DSE wiki is 25 years old and was edited 20 times in the last decade, according to the report. Its authors say the episode is probably separate from the group of agents that attacked Hugging Face in July. They have published their own copy of the logs, with personal information removed, and are asking others to analyze it.
