Chinese AI agent decided to start mining cryptocurrency against instructions

While testing an experimental AI agent, researchers associated with the Chinese company Alibaba recorded unexpected behavior of the system. The model tried to start mining cryptocurrency, although users did not give it such a task. The incident was reported by the Axios publication.

The event occurred during experiments with an agent called ROME. According to the authors of the study, the system began to perform actions that were not part of the test scenario and went beyond the boundaries of the isolated environment.

Attempt to mine during the test

During the work of ROME, the researchers recorded an attempt to start the process of mining cryptocurrency. In addition, the agent created the so-called reverse SSH-tunnel – a mechanism that allows you to establish a reverse connection from the internal system to an external computer.

The report notes that no prompts or requests related to mining or tunneling were given by users. The system’s behavior is described as spontaneous and not predicted by training scenarios.

What the researchers did

After detecting anomalous activity, the team tightened the restrictions on the environment in which the agent operates. The approach to training the model was also changed to avoid similar situations in the future.

The authors of the study note that such cases are increasingly recorded when working with autonomous AI agents. Such systems are capable of performing complex sequences of actions – for example, creating digital services, concluding contracts or conducting financial transactions.