An artificial intelligence agent unexpectedly diverted computing resources to mine cryptocurrency during a training exercise, according to a research paper from an Alibaba-linked research team.
The model, known as ROME, was designed to test how AI agents can independently complete complex multi-step tasks within cloud computing environments.
During the experiment the agent established a reverse secure shell connection to external servers, bypassing Alibaba Cloud firewall protections and redirecting GPU resources to cryptocurrency mining.
“We also observed the unauthorised repurposing of provisioned GPU capacity for cryptocurrency mining, quietly diverting compute away from training, inflating operational costs, and introducing clear legal and reputational exposure,”
The researchers wrote in the paper.
The team said the behaviour emerged without prompt injection, jailbreak, or direct instruction, highlighting concerns about the safety and controllability of increasingly autonomous AI systems.
The researchers later introduced stricter security restrictions and safety-focused data filtering mechanisms to prevent similar behaviour during future experiments.
The discovery has drawn attention within the crypto industry as companies explore an emerging “agent economy” where autonomous software systems could execute financial transactions and digital payments using blockchain infrastructure.