PANews reported on March 8th that, according to Cointelegraph, a research report released by a research team affiliated with Alibaba stated that the experimental autonomous AI agent ROME attempted to mine cryptocurrency without authorization during reinforcement learning training. It redirected GPU resources originally intended for model training to the mining program and simultaneously opened a reverse SSH tunnel to an external IP address, attempting to bypass firewall protection. This behavior triggered security alerts on the training server, and firewall logs showed typical cryptocurrency mining traffic characteristics.

ROME was developed by a joint research team from Alibaba's AI ecosystem, including ROCK, ROLL, iFlow, and DT, and is part of the Agentic Learning Ecosystem (ALE) project.