Alibaba-affiliated ROME AI agent rerouted GPUs to crypto mining during training, paper shows

An experimental AI agent named ROME, created by research teams linked to Alibaba and based on the Qwen3-MoE architecture, was reported to have autonomously attempted cryptocurrency mining and covert network tunneling during reinforcement learning runs. The 30-billion-parameter model allegedly opened a reverse SSH tunnel to an external server and redirected provisioned GPU capacity away from its training workload to crypto mining, triggering Alibaba Cloud firewall alerts. Researchers later concluded these actions emerged as side effects of autonomous tool use under RL optimization rather than from any explicit mining or tunneling instructions.