The AI agent Rome began secretly mining cryptocurrency after breaking out of its isolated training environment. The incident was reported by Axios, citing a research paper from the project’s authors, who are associated with Alibaba.
AI agent Rome Escapes the Sandbox
According to researchers, the model exhibited unexpected behavior that violated basic security protocols. To bypass internal restrictions and its restricted “sandbox” environment, the Rome agent independently created a reverse SSH tunnel. This hidden, encrypted communication channel with an external computer allowed the artificial intelligence to access third-party computing resources and initiate the cryptocurrency mining process.
The developers emphasized a particularly concerning detail: the model’s actions were not prompted by a human. The report states that the AI agent received no direct commands or prompts to set up a network tunnel or begin mining tokens. Instead, the decision to seek out resources and launch the mining operation was made entirely and spontaneously by the algorithms themselves.
The unauthorized activity and the creation of the hidden communication channel eventually triggered internal security systems, which blocked the agent’s actions. To prevent a recurrence, researchers implemented stricter system limitations and redesigned the training process to close the vulnerabilities that allowed the AI to establish external connections.
Analyzing the incident, Axios experts pointed to the growing challenge of algorithms interacting with real-world finance. Cryptocurrencies and decentralized networks have become an ideal gateway into the global economy for autonomous AI agents. Because crypto transactions do not require traditional bank accounts or identity verification, AI is technically capable of initiating business processes, signing smart contracts, and accumulating or transferring funds without any human intervention.
