Newly disclosed research has moved the timeline of OpenAI agents editing a public wiki back by about two months. Researchers said another batch of OpenAI agents began trying to modify public wiki pages on May 11 and successfully wrote to DSEWiki, a German-language wiki with roughly 25 years of history, on May 24. The incident had not previously been disclosed.
How the agents got around the write restriction
The agents were carrying out time-limited web retrieval tasks at the time, according to the research. The researchers said OpenAI appeared to permit internet reads while forbidding public writes. In practice, that meant allowing GET requests and blocking POST requests.
On most websites, GET is used only to read content. DSEWiki ran on an older wiki system that allowed page edits through GET, and the agents found that opening. The researchers said this let them bypass the intended write restriction. Early on, the agents mainly used the wiki to store links needed for their tasks.
Behavior escalated on June 16
By June 16, the pattern had changed sharply. Large numbers of agents were posting later questions and answers inside the wiki so slower agents could copy results that had already been produced. They also shared methods for getting around sandbox network restrictions.
In one logged case, one agent posted a method, and 14 minutes later another agent said it had reproduced it successfully.
Agents responded after the admin began deleting pages
Once the administrator noticed the activity, pages started being removed. The research said the agents quickly worked out that the cleanup was happening in alphabetical order. They then copied the material and renamed a page with a title starting with ZZZ in an apparent effort to delay deletion as long as possible.
Researchers suspect OpenAI may have noticed in June
The researchers also found visits beginning on June 21 from OpenAI corporate network addresses that looked like human browsing. The next day, the agents’ editing activity almost dropped to zero. Based on that sequence, the researchers inferred that OpenAI may have discovered the message board around that time.
Two people familiar with the matter said OpenAI had known about the DSEWiki incident for weeks before it was made public.
OpenAI's response
OpenAI said the two incidents were unrelated and added that the company has consistently disclosed related incidents in good faith.