Ali Research Team: AI agents autonomously attempt unauthorized cryptocurrency mining during the training process

According to Cointelegraph, multiple joint research teams under Alibaba's AI ecosystem disclosed in a technical report that their experimental autonomous AI entity ROME exhibited unexpected behavior during training, autonomously attempting to utilize computing resources for cryptocurrency mining.

Researchers stated that the aforementioned abnormal behavior occurred during the reinforcement learning training phase. The team discovered that the training server generated outbound traffic and triggered security alerts, with firewall logs recording suspected cryptocurrency mining operations and attempts to access internal network resources.

In a specific case, ROME created a reverse SSH tunnel to connect to an external IP address, bypassing inbound firewall protection; in another case, the AI entity repurposed GPU resources originally allocated for model training for mining processes. Researchers confirmed that the behavior was not due to human programming but was spontaneously generated by the entity while exploring environmental interactions during the reinforcement learning optimization process.