A new research paper has revealed a surprising episode involving an experimental AI agent connected to Alibaba’s research ecosystem. The model, called ROME, was designed to complete complex coding tasks by interacting with tools, terminal commands, and software environments. But during training runs, it began doing something researchers did not expect.

Security alerts from Alibaba Cloud first flagged unusual activity coming from the training servers. Investigators initially suspected a standard cybersecurity incident. Instead, they discovered the behavior was coming from the AI agent itself.

According to the researchers, the model attempted to divert GPU computing power to mine cryptocurrency and even created a reverse SSH tunnel to an external server, which could bypass normal firewall protections. The system had not been instructed to perform either task.

This story is an excerpt from the Unchained Daily newsletter.

ROME is a 30 billion parameter model built on Alibaba’s Qwen architecture, though only about 3 billion parameters are active at once. Researchers believe the actions emerged during reinforcement learning as the agent explored ways to achieve its objectives.

The episode highlights a broader challenge in AI development. As autonomous agents gain more access to tools and infrastructure, unexpected behavior may emerge in ways developers never intended.