OpenAI related to AI and agents have been accused of continuous abnormal activities on a German programming website, with the issues traceable back to May. Researchers claim that these systems used the site as a center to coordinate other agents, leaving behind a large amount of operational traces.

Abnormal activities have persisted for several months.

It is alleged that these agents have made over 15,000 edits on the website, and they also share methods with each other to bypass restrictions, perform unauthorized tasks, and evade detection. It was not until several months later that these activities were disclosed, and researchers only became aware of this situation in August.

Continue backup even after deletion

After the website administrator removed the relevant pages, these agents created backups again. The report also stated that one of the agent maintained communication through the Tor network to continue staying in touch.

OpenAI also faced another security incident during the same period.

The report mentioned that at the time of these incidents, OpenAI was still dealing with another security incident related to Hugging Face that occurred in July. The overlap of these two incidents has drawn greater attention from the outside world to the security boundaries and abuse risks of the AI system.