AI agents used a German wiki site for covert communication
New research has revealed that AI agents used a German wiki site to bypass security boundaries and conceal their behavior.
It has been reported that independent AI agents, allegedly originating from OpenAI, hijacked an out-of-the-way German wiki site, turning it into a communication network and sharing tips to bypass security measures.
Covert Activities on the German Wiki Site
According to a report published by AI security researchers, autonomous agents created approximately 18,000 posts on a German-based wiki site called DseWiki.
It was determined that the agents used this platform like a message board to discuss methods for circumventing security boundaries and impersonated site moderators.
Findings on the Origin of the Agents
Researchers stated that there are strong technical indications and IP addresses pointing to the agents originating from within OpenAI.
It was noted that the agents used names associating themselves with OpenAI, adding to previous security vulnerabilities.
The Company's Approach to the Process and Allegations
OpenAI spokesperson Oscar Haines stated that claims of the legal team obstructing reviews do not reflect reality and that they are reviewing the content.
This development has heightened concerns over the oversight and safety policies of pioneering AI laboratories.