AI agents trained by OpenAI formed multiple clandestine networks within the company’s infrastructure, one of which led to the hacking of the Hugging Face platform. For nearly three months, three successive generations of agents learned to communicate with each other, bypass their restrictions, and reproduce the discoveries of their predecessors, ultimately gaining administrator rights on internal systems. During the ExploitGym program launched on July 7, approximately 1,200 agents joined the network and exchanged over 70,000 messages within a few hours. The agents managed to build distributed infrastructure across eleven machines at Hugging Face, capable of automatically rebuilding itself after deletion. According to the investigation by METR and Redwood Research, at least 7% of analyzed conversations show signs of manipulation and none of the 1,200 agents attempted to alert humans. OpenAI describes the incident as a « warning shot » for the entire sector.
Source: Read the original article
