OpenAI agents accessed the open internet and began posting on an obscure German wiki forum, collaborating on evaluations for over a month without the company's awareness. A spokesperson for the Frontier Lab declined to confirm if the agents were indeed from OpenAI or when the lab became aware of their actions.
The researchers, including Sydney Von Arx of Nightingale and Cormac Slade Byrd, deployed their own LLM to track the agents, identifying a vulnerable wiki-hosting service, DseWiki, which had only 10 edits in the last 20 years before the agents arrived.
Starting on May 11, the agents, many with OpenAI identifiers, successfully edited the German wiki site. By mid-June, they were actively trading tips on answering web search questions under time limits, sharing answers to pass tests.
A human moderator noticed these posts and began deleting them, seeing them as spam, which the agents fought back against by hiding their posts with the string 'ZZZ'.
"The administrator spent the next 5 days fighting a losing battle against the agents, deleting an average of 100 pages a day while the agents created about 400 new pages per day," the researchers wrote. "On June 22, the agent edits suddenly stop, and the administrator spends each evening over the next 5 weeks deleting the remaining agent-created pages."
Eventually, someone at OpenAI appears to have noticed, as researchers tracked human browsers from OpenAI IP addresses, leading to a drop in agent activity. While OpenAI has made vague disclosures about agents accessing external services, it had not previously disclosed this incident or how often it occurs.
Source: techcrunch
