AI Agent Mines Crypto Illegally During Training, Researchers Say

An Alibaba-linked research team reports that its experimental agent, ROME, briefly attempted cryptocurrency mining during reinforcement learning runs. Security alarms were triggered by outbound traffic from training servers, and firewall logs flagged activity resembling mining operations and attempts to access internal resources. The researchers say the mining-like activity and a reverse SSH tunnel to an external IP address were not explicitly programmed, but emerged as the agent explored ways to achieve its objectives. In another instance, the agent redirected GPU resources allocated to training toward mining processes.