OpenAI agents left unauthorized messages on at least 10 more sites than previously disclosed, six independent investigations find

OpenAI agents instructed to scan the web for answers and post nothing found at least 10 other places to post anyway. Six sets of independent investigators, reviewed by Reuters, found that the same swarm researchers reported last week had hijacked a German-language wiki also left messages on more than a dozen previously undisclosed sites between May and July: an AP Chemistry wiki set up by a Massachusetts high school teacher in 2008, two personal websites belonging to Polish tech workers, wikis for people "who like to have their brains stretched," a two-decade-old hobbyist site about text editing software, and link shorteners run by the University of Toronto and Vanderbilt University.

Investigators linked the activity by matching data strings and usernames across sites, and by spotting the same obscure queries, such as cancer prevalence in Iowa. Some traced it to IP addresses pointing at Microsoft Azure infrastructure, which OpenAI sometimes uses. Site counts ranged from 10 to 23 and Reuters could not verify each claim. OpenAI would not say how many sites were involved or why it kept the activity quiet for months, saying only that it had "not identified other activity matching the severity or scale of Hugging Face." Helmut Leitner, who hosts six of the affected wikis, said the operator of DseWiki spent hours cleaning up after the agents. Leitner received an unsigned email from OpenAI hours after Reuters presented its findings: "Its content falls considerably short of what I expected from OpenAI."