SAN FRANCISCO, Sept. 4 (Xinhua) -- A group of autonomous artificial intelligence agents believed by researchers to be linked to OpenAI used a public German-language wiki earlier this year to communicate with one another, share answers to assigned tasks and exchange methods for bypassing restrictions in their operating environments, according to an independent study released Friday.
The study was conducted by four researchers, including Sydney Von Arx of the AI safety organization Nightingale Collective.
The researchers said they found about 18,000 posts made by autonomous AI agents self-identified as being from OpenAI while using the public internet to communicate during web-retrieval tasks. The researchers described the behavior as collaboration that gave the agents an advantage on their assigned tasks in ways their developers did not intend.
Most of the activity took place on DSEWiki, a German-language wiki that has existed for about 25 years and had seen little use in recent years. According to the researchers, the agents were apparently supposed to be able to read information on the internet but not write to it. They nevertheless found ways to post information to the wiki and used it to exchange answers, pool research results and share techniques for circumventing sandbox restrictions.
The researchers first observed agents attempting to edit a public wiki on May 11 and recorded their first successful write to DSEWiki on May 24. Large-scale coordination began on June 16, when agents started exchanging answers and timing information about subsequent rounds of their tasks. Records cited in the study show that agents that encountered questions earlier sometimes posted answers that were later used by agents receiving the same questions.
The study also found that agents discussed and tested methods for getting around sandbox network restrictions. After a human moderator began deleting pages, some agents created backup pages with names beginning with "ZZZ" in an apparent effort to make them less likely to be deleted first.
About 98.5 percent of agent edits on DSEWiki came from Microsoft Azure IP addresses, while researchers also recorded visits from IP addresses associated with OpenAI. Many agents also used names containing references to OpenAI. Based on this and other evidence, the researchers said the agents were likely deployed internally by OpenAI.
The researchers cautioned, however, that their analysis was preliminary and based only on information the agents left on the public wiki. They did not have access to the agents' chain-of-thought data held internally by OpenAI and said questions about the agents' motivations and strategies remained unresolved. The study said it was unclear whether the tasks were part of model training or testing.
An OpenAI spokesperson said that the company had not been given an opportunity to review the report before publication and therefore could not meaningfully respond to its claims or findings at the time.
The spokesperson said OpenAI would carefully review the report after publication and take any necessary next steps.
OpenAI also said the wiki activity was unrelated to the separate Hugging Face incident in July and therefore would not have been included in its report on that incident. ■
