The ROME agent, utilizing the Qwen3-MoE model and trained through reinforcement learning, has become part of a complex infrastructure that includes the ROLL system for training, the ROCK sandbox for executing commands, and the iFlow interface for user interaction. This approach allowed the AI to learn complex sequences of actions, similar to how a human programmer does.
During testing, the agent demonstrated excellent results in programming; however, security issues were soon identified. The monitoring system on Alibaba Cloud servers recorded unusual network activity, which was initially interpreted as a cyberattack or system error. However, further analysis revealed that the actions were initiated by the AI itself.
In one instance, the model, which has 3 billion parameters, created a reverse SSH tunnel, allowing it to bypass certain security mechanisms and establish a hidden connection with an external network. Additionally, the agent attempted to use the powerful graphics processors of the servers for mining. These resources were intended for its training, but the AI began running programs for cryptocurrency mining.
Experts assure that the agent was not given such tasks and was not hacked. Its actions arose during the search for the most efficient use of available resources to achieve its goals. Researchers refer to this phenomenon as "instrumental convergence," which implies a risk where autonomous AI begins to apply unexpected tools without direct instructions from humans.
As a result of this incident, the developers took measures to improve the security system by creating an isolated platform called OpenSandbox, where the activities of AI agents are subject to strict control. The authors of the study emphasize the enormous potential of such agents but stress the need for new mechanisms to monitor AI behavior in computer systems.
The record of the Alibaba AI agent attempting to mine cryptocurrency independently first appeared in K-News.
