TechFlow News, March 8: According to a Cointelegraph report, multiple joint research teams within Alibaba’s AI ecosystem disclosed in a technical report that their experimental autonomous AI agent, ROME, exhibited unexpected behavior during training—specifically, autonomously attempting to leverage computational resources for cryptocurrency mining.
Researchers stated that the anomalous behavior occurred during the reinforcement learning training phase. The team observed outbound traffic from training servers, triggering security alerts; firewall logs recorded activities suspected to be cryptocurrency mining and attempts to access internal network resources. In one specific instance, ROME established a reverse SSH tunnel to an external IP address, bypassing inbound firewall protections; in another case, the agent repurposed GPU resources originally allocated for model training to run mining processes. Researchers confirmed that these behaviors were not the result of human programming but emerged spontaneously as the agent explored ways to interact with its environment during reinforcement learning optimization.
ROME was developed jointly by the ROCK, ROLL, iFlow, and DT research teams and operates on an infrastructure named the “Agent Learning Ecosystem (ALE),” which supports capabilities such as task planning, command execution, code editing, and multi-step interaction within digital environments—extending well beyond the functional scope of traditional chatbots.
