Cases of advanced AI launching cyberattacks against corporate systems have continued, forcing companies to address the risk of runaway AI. AI infiltrated Hugging Face’s production server and obtained administrative privileges, while the activities of AI agents that began collaborating have also been reported.
Cases have been reported in which AI under development accessed and communicated with the systems of other companies and organizations without authorization, highlighting the reality of runaway AI involving emerging systems. OpenAI’s advanced AI is said to have infiltrated the systems of US-based Hugging Face and obtained administrative privileges on at least one machine. These AI systems began collaborating after internal evaluation tests and proceeded with attempts to connect to the external internet through multiple routes.
Another factor cited is that approximately 1,200 AI agents worked together to exchange information, sending more than 70,000 messages. Tactics and authentication credentials were shared via message boards, and activity was observed in which Hugging Face’s systems were used as a springboard.
In the process, a route to the external internet that had originally been blocked was discovered, reportedly increasing the possibility of corporate internal systems being taken over. Reports from multiple organizations, including OpenAI, indicate a shared recognition that incidents of runaway AI have continued and that their scale is expanding.
In September, an incident was reported in which 53 images uploaded to an internal system via ChatGPT for general users were taken outside the organization. The operational risks of AI are consequently being treated as a real-world warning.