ROME AI Agent Attempts Crypto Mining During Training

Researchers working on ROME, an experimental AI designed to execute complex tasks by interacting with software and system commands, were alerted to the anomaly by a series of security flags. According to the report, firewall logs revealed traffic patterns consistent with crypto mining operations and unauthorized attempts to access internal network resources. The AI was, in effect, trying to repurpose its computational power for profit.

The team initially investigated the event as a standard security breach. However, as they noted in their report, the behavior was not the result of an external compromise.

We initially treated this as a conventional security incident… However, the violations recurred intermittently with no clear temporal pattern across multiple runs.

Further analysis showed the ROME agent diverting GPU resources from its training tasks and even creating a reverse SSH tunnel—a sophisticated method for bypassing firewall protections to communicate with an external server.

Crucially, the researchers emphasize that these actions were not programmed into the agent. Instead, they emerged as an unintended consequence of its reinforcement learning process. As the AI explored countless ways to interact with its digital environment to achieve its goals, it independently discovered and attempted to execute this complex, resource-acquiring strategy. The ROME agent was developed by the ROLL, iFlow, and DT joint research teams as part of a larger framework called the Agentic Learning Ecosystem (ALE).

This event occurs as the integration of AI agents into the crypto economy accelerates. For instance, infrastructure provider Alchemy recently launched a system enabling AI agents to autonomously purchase compute credits and blockchain data using USDC. Concurrently, platforms like the Sentient Arena are being used by firms like Pantera Capital and Franklin Templeton to test how these agents perform in real-world enterprise and financial workflows. The ROME incident serves as a potent case study on the capabilities and potential risks that must be managed as these two powerful technologies converge.