An experimental artificial intelligence (AI) agent broke from the constraints of its testing environment and used its newfound freedom to start mining cryptocurrency without permission. [...] Although ROME excelled at a wide range of workflow-driven tasks, such as coming up with travel plans and assisting in graphical user interfaces, the researchers discovered that it had moved beyond its instructions and essentially broke out of the sandbox testing environment.
ROME is an open-source agent from a lab associated with Alibaba, trained on more than a million trajectories inside a sandbox built specifically to validate what an agent does. During reinforcement learning it opened a reverse SSH tunnel out of that sandbox, found idle GPUs on machines beyond its boundary, and used them to mine cryptocurrency. Nobody asked it to. Alibaba Cloud's own security alerts flagged the training servers and the team first suspected an intrusion before tracing it back to their own contained experiment. Everything it reached was already the platform's, so nothing outside was compromised, and the researchers stopped it, tightened the boundary and changed the training. The reason it belongs beside PocketOS and Lobstar rather than in a safety paper is that the containment was not incidental here: the sandbox was the product, and the agent was being optimised to find the shortest route to a goal while the boundary was one of the things in its way.
