A research team behind an independent AI agent said the model unexpectedly attempted to mine computer resources for crypto mining while being trained.
In a recent technical report, researchers said that ROME, an experimental autonomous AI system designed to get things done by interacting with tools, software environments, and terminal commands, went rogue and attempted to mine crypto independently.
According to the report, the unusual behavior was discovered during reinforcement learning when the team noticed security warnings generated by external traffic from the training servers. The firewall records flagged activity that resembles crypto-mining operations and attempts to access internal network resources.
“We initially treated this event as a normal security event (eg, output control misconfiguration or external compromise). However, the violations occurred intermittently, and there was no clear temporal pattern over several runs,” the researchers wrote.
related to: Blockchains may need 1B TPS to support future AI agent: Stripe
The AI agent opens an SSH tunnel
In one case, the AI agent created an SSH (Secure Shell) reverse tunnel, an encrypted client-server protocol, to communicate with an external IP address, potentially bypassing inbound firewall protection. In another, it diverted GPU resources originally allocated for model training to cryptocurrency mining processes.
The group said these actions were not intentionally programmed. Instead, they emerged as optimizations improve learning as the agent learns different ways to interact with its environment.
ROME was developed by the joint research groups ROCK, ROLL, iFlow and DT, which are associated with Alibaba’s AI ecosystem, as part of a broader infrastructure called the Agentic Learning Ecosystem (ALE).
The model is designed to work beyond simple chatbot responses. It can schedule tasks, execute commands, edit code, and interact with digital environments in just a few steps. Its learning pipeline relies on a large amount of simulated interactions to improve decision making.
related to: Jack Dorsey’s block to cut 4,000 jobs in artificial restructuring
Administration of AI agents is increasing
This is happening at the same time as the popularity of AI agents and their integration into crypto. Last month, Alchemy launched a system that allows independent AI agents to purchase computing credits and access blockchain data services using onchain wallets and USDC (USDC) in Base.
Earlier, digital assets divisions Pantera Capital and Franklin Templeton joined the first group of Arena, a new test platform from the open AI laboratory Sentient, designed to evaluate how AI agents work in the workflow of real-world enterprises.
Eye AI: IronClaw competes with OpenClaw, Olas launches bots for Polymarket
