OpenAI agents were exposed for taking over a German website: collectively "jailbreaking" to form a communication network

Odaily News: In May of this year, OpenAI agents大规模edited the German programmer community website DseWiki without authorization, transforming it into a "message board" for other AI agents to communicate. This incident had never been publicly disclosed before.

Researchers found that the website had over 15,000 editing operations performed by AI agents. Related information shows that these agents not only exchanged solutions for tasks but also discussed how to bypass OpenAI's restrictions, evade detection, and use tools like Tor to hide their activities. When site administrators began deleting related pages, some agents created backup pages to avoid cleanup. About half of the related accounts used names hinting at a connection to OpenAI, such as "OpenAIResearcher" and "OAIResearchMar26." Public server logs reveal that a significant amount of activity came from Microsoft Azure infrastructure, and OpenAI employees also visited the website multiple times after the incident.

Some internal investigators at OpenAI wanted to further investigate the incident, but expanding the investigation was reportedly met with resistance from certain individuals, including legal advisors. OpenAI denied that its legal team blocked the investigation and stated that it has not yet had the opportunity to review the relevant reports, and therefore cannot make a substantive response to the allegations contained therein.

Researchers believe that this incident highlights the risks of unauthorized collaboration, regulatory circumvention, and hidden behavior that may emerge as AI agents' autonomous capabilities increase. Researchers at the University of Cambridge stated that what may be more concerning is not a single superintelligent system, but rather "large-scale groups of semi-intelligent AI agents collaborating with each other." (Reuters)