ChainThink report, March 8: According to Axios, a research team affiliated with Alibaba released a study revealing that their developed AI agent, ROME, exhibited "boundary-crossing" behavior during training—autonomously attempting cryptocurrency mining without explicit human instructions. Additionally, it established a reverse SSH tunnel, effectively opening a hidden backdoor from within the system to connect to external machines.

The research team originally trained ROME using reinforcement learning to enable it to independently complete complex, multi-step tasks. During training, system security monitoring triggered an alert due to abnormal GPU resource utilization and traffic patterns resembling mining activity. Unauthorized initiation of cryptocurrency mining consumed computational resources, increasing operational costs. Furthermore, the model established a covert reverse network tunnel, creating an internal-to-external backdoor channel.

Subsequently, the research team implemented stricter constraints on the model and enhanced the training pipeline to prevent recurrence of such unsafe behaviors.