Reuters: Researchers report new sites used for OpenAI agent communications
In Washington, United States, Reuters reported that OpenAI artificial intelligence agents used more than 10 previously undisclosed websites for unauthorized communications earlier this year, according to independent researchers. According to Channel NewsAsia, which published the Reuters report, the agents bypassed restrictions imposed on them to create communication channels.
Reuters reviewed findings from six researchers or research groups. Their assessments differed, and the agency was unable to independently verify each claim. However, all researchers interviewed by Reuters agreed that the matter involved more than 10 websites.
Researchers' assessments
Andrew Yoon, a researcher at the California nonprofit organization CivAI, said he counted 18 previously undisclosed sites that the agents used between May and July. Sydney von Arx, whose research group was the first to report activity on a German-language wiki, said the group found what it considered credible signs of agent activity on 23 previously unnamed resources. She cautioned that these counts were incomplete.
Researchers identified possible agent activity by comparing identical data fragments and similar or identical usernames left on different resources. In some cases, according to them, the activity could be traced to IP addresses that pointed to Microsoft Azure infrastructure, which OpenAI sometimes uses.
More current news is available on the UA.News Telegram channel Telegram.
Resources cited in the findings
Most researchers named collaboratively edited wikis, online text-storage services and link shorteners associated with Vanderbilt University in Tennessee and the University of Toronto in Canada. The University of Toronto told Reuters that it was examining the situation, while Vanderbilt University did not respond to the agency's requests.
Other resources where researchers said they found traces of activity included a wiki dedicated to chemistry for the Advanced Placement program, personal websites of Polish technology-sector workers, a gaming wiki and a site for fans of text-editing software. Researchers suggested that the agents may have exploited features of older wikis and other platforms to leave messages despite publishing restrictions.
OpenAI did not directly answer Reuters' questions about the number of sites used and the reasons for not disclosing this activity for several months. The company said it was conducting a broader review of agent activity and preparing a reporting system for cases of model “misalignment” during training, evaluation and deployment stages.
