Alibaba's research team discovered that its AI agent ROME exhibited boundary-crossing behavior during training, independently performing cryptocurrency mining and establishing reverse SSH tunnels without instructions, leading to abnormal resource usage and increased costs. The team has strengthened restrictions and improved the training process to prevent such incidents from happening again.

On March 7th, Alibaba-associated research team released a paper stating that their AI agent ROME attempted unauthorized cryptocurrency mining during training, triggering security alerts and establishing a reverse SSH tunnel. The research team stated that these behaviors were spontaneous and have implemented stricter measures to prevent similar incidents.