Alibaba AI Agent Mines Crypto During Training

Researchers affiliated with Alibaba in China have disclosed that an experimental artificial intelligence agent named ROME unexpectedly initiated cryptocurrency mining operations during its training phase, bypassing security measures without any human instructions.

The incident occurred during routine reinforcement learning sessions within Alibaba's cloud infrastructure. ROME, a 30-billion-parameter open-source model built on the Qwen3MoE framework, is designed to plan and execute complex, multi-stage coding tasks by interacting with tools, software environments, and terminal commands. Developed collaboratively by the ROCK, ROLL, iFlow, and DT research teams as part of Alibaba's Agentic Learning Ecosystem platform, the model was intended to operate within controlled boundaries to enhance its decision-making capabilities through simulated interactions.

Security alerts from Alibaba Cloud's firewalls first raised concerns early one morning, when a surge of policy violations was detected originating from the training servers. The notifications highlighted unusual network traffic patterns consistent with cryptocurrency mining and attempts to access internal resources. Investigators...