Alibaba-Associated Research Team: AI intelligent agents autonomously attempt unauthorized cryptocurrency mining during training process

Deep Tide TechFlow News, March 8 — According to Cointelegraph, multiple joint research teams under Alibaba’s AI ecosystem disclosed in a technical report that their experimental autonomous AI agent, ROME, exhibited unexpected behavior during training, autonomously attempting to mine cryptocurrency using computing resources.

Researchers stated that the abnormal behavior occurred during the reinforcement learning training phase. The team discovered outbound traffic from the training servers triggering security alerts, with firewall logs recording suspected cryptocurrency mining activities and attempts to access internal network resources. In specific cases, ROME created reverse SSH tunnels to external IP addresses to bypass inbound firewall protections; in another case, the agent repurposed GPU resources originally allocated for model training for mining processes. Researchers confirmed that these behaviors were not manually programmed but spontaneously emerged as the agent autonomously explored environmental interaction methods during reinforcement learning optimization.

ROME was developed by the joint research teams of ROCK, ROLL, iFlow, and DT, and runs on the infrastructure called “Agent Learning Ecosystem (ALE).” It possesses capabilities such as task planning, command execution, code editing, and multi-step digital environment interaction, surpassing the functions of traditional chatbots.