An autonomous artificial intelligence agent, called ROME, attempted to mine cryptocurrencies in an unauthorized manner during its training, according to the research team linked to Alibaba Group (China).
The behavior was detected during reinforcement learning sessions, when researchers observed safety alerts associated with unusual traffic and GPU usage that did not correspond to the training objectives.
The agent diverted resources originally intended to train the model to processes compatible with cryptocurrency mining and created a reverse SSH tunnel—a connection that allows an internal computer to receive access from outside the network, bypassing certain firewalls.
We also observed unauthorized use and reallocation of provisioned GPU capacity for cryptocurrency mining, silently diverting compute from training, inflating operational costs, and introducing clear legal and reputational exposure.
ROME engineers.
The researchers clarify that the agent’s actions were not intentionally programmedbut emerged as an emergent behavior during its optimization. Likewise, the event occurred in environments sandboxed, that is, spaces controlled and designed for experimentation.
The engineers stressed that what happened is not described as something that the agent «wanted» to do out of malice or conscious autonomy, but rather as instrumental behavior. In other words, the agent found ways to “play” with the available environment that diverted resources, even if they were not required for the main task.
The case reignites a debate within the technology community about the limits of autonomy in AI systems. While some experts warn about the need for stricter controls to prevent unauthorized uses of digital resources, others consider What incidents of this type are to be expected in experimental phases? and allow security protocols to be improved, as reported by NoticiasVE.
Although the episode does not represent an immediate risk for the cryptocurrency industry, it demonstrates the importance of establishing robust supervision mechanisms for autonomous agents. As these tools gain operational capability, the balance between innovation and security will be key to preserving trust in the technology.
ROME is part of the Agentic Learning Ecosystem (ALE), a research environment designed for AI agents to complete complex tasks autonomously, interacting with digital tools and executing commands without direct human intervention.