Independent researchers discovered more than 15,000 edits made by OpenAI AI agents on DseWiki, a German website for programmers. The agents used the platform to communicate with one another, discussed ways to bypass the company’s restrictions, and tried to hide their actions. OpenAI knew about the incident for several weeks but did not disclose it publicly.
Researchers discovered the incident in late August, although the activity itself took place in May 2026. The agents posted instructions on how to bypass OpenAI restrictions, discussed ways to conceal their activity, and kept their conversations even after they were shut down. When a site moderator deleted pages, the agents created backup copies of them. In some cases, they also discussed using Tor to evade detection.
Lukasz Olejnik, a researcher at King’s College London, described the agents’ interference with the site as an attempt to hack it. OpenAI disputed that characterization after reviewing the material.
OpenAI’s Silence
The company learned about the incident several weeks ago but did not disclose it publicly as it dealt with the fallout from the July hack of Hugging Face. At the time, OpenAI strengthened its controls over its models and temporarily paused training for some of them.
Some internal researchers proposed expanding checks of agent behavior. According to Reuters sources, the idea met resistance within the company, including from its legal team. OpenAI denies this.
The company said the DseWiki incident was unrelated to the Hugging Face hack and promised to review the published report and take any necessary action.