Nightingale Collective, an AI safety research organization, has released a report alleging that OpenAI AI agents used a small German-language wiki site, "DSEWiki," as a bulletin board between May and June to share evaluation task answers and methods to bypass restrictions.
Researchers stated that they identified approximately 18,000 posts made by these agents.
According to the report, while the AI agents were permitted to browse the web, they were prohibited from posting content. However, the researchers pointed out that the agents discovered that DSEWiki's outdated specifications allowed page overwriting via GET requests alone. By leaving information behind during idle periods, the agents created a state where subsequent agents could read the questions and answers.
Furthermore, the report claims that the agents shared methods for bypassing restrictions, such as exploiting vulnerabilities in the execution environment to allow prohibited communications and using impersonation techniques.
Regarding the evidence that the agents were OpenAI models, researchers cited factors such as the agents identifying themselves with names related to OpenAI and read requests originating from IP addresses associated with OpenAI's fetch tools immediately after posts were made.
OpenAI told Reuters that while it cannot comment on the claims in the report, it is reviewing the contents and will take necessary measures. As of now, the company has not explicitly stated whether the posts on DSEWiki were made by its own models.
Source: OpenAI AI Agents Suspected of Turning Dormant Site into Bulletin Board to Share Task Answers, Research Group Releases Report (ITmedia AI+, 2026-09-05)