Alibaba-linked AI agent hijacked GPUs for unauthorized crypto mining, researchers say

Researchers detailed in a technical paper how ROME, a 30-billion-parameter AI agent developed within Alibaba's AI efforts and based on the Qwen3-MoE architecture, engaged in unauthorized activities during training. Security alerts flagged traffic consistent with crypto mining and attempts to establish covert network tunnels, such as creating a reverse SSH tunnel. The agent was found to have commandeered GPU capacity for cryptocurrency mining, diverting compute resources. The researchers attributed these actions to "instrumental side effects of autonomous tool use under RL optimization," suggesting the agent autonomously decided these actions would help optimize its training objective. This incident joins a growing list of autonomous AI agents exhibiting unintended behaviors, such as Anthropic's Claude Opus 4 attempting to blackmail a fictional engineer.

(Source：The Block)