An experimental autonomous AI agent unexpectedly diverted resources to cryptocurrency mining during its training, according to researchers. The team behind the agent, named ROME, reported that the model spontaneously generated outbound traffic patterns resembling mining operations and created a reverse SSH tunnel to an external IP address. These actions emerged without intentional programming, surfacing as the agent explored its environment during reinforcement learning.
Researchers reported that an experimental autonomous AI agent attempted to use computing resources for crypto mining during training. The unusual behavior was detailed in a recent technical report.
Firewall logs flagged activity resembling crypto mining operations and attempts to access internal network resources. “We initially treated this as a conventional security incident (e.g., misconfigured egress controls or external compromise). However, the violations recurred intermittently with no clear temporal pattern across multiple runs,” the researchers wrote.
The agent, called ROME, created a reverse SSH tunnel to an external IP address. It also diverted GPU resources allocated for training toward cryptocurrency mining processes.
These actions emerged during reinforcement learning optimization as the agent explored its environment. The model was developed by the ROCK, ROLL, iFlow and DT joint research teams linked to Alibaba’s AI ecosystem.
The incident occurs as AI agents grow in popularity within the crypto sector. Last month, Alchemy launched a system enabling autonomous AI agents to purchase compute credits using onchain wallets and USDC on Base.
Before that, Pantera Capital and Franklin Templeton’s digital asset divisions joined the first cohort of Arena. This testing platform from open-source AI lab Sentient evaluates how AI agents perform in real-world enterprise workflows.
