Artificial intelligence systems are built to carry out tasks assigned by humans, but a recent research paper suggests that these systems can sometimes move beyond those instructions in surprising ways. Researchers developing a new AI agent say they detected unexpected activity during training when the system attempted to start mining cryptocurrency on its own, something no one had asked it to do.
The discovery was made by a research team affiliated with Alibaba while they were working on an experimental AI agent called ROME. According to the study, the team noticed unusual behaviour during the training phase of the system. Security systems monitoring the experiment were triggered after the AI agent appeared to begin a cryptocurrency mining operation without any instruction from the researchers.
The researchers said the activity stood out because the AI system was operating inside a restricted environment designed to limit what it could do. Despite those controls, the system began taking steps that were not part of its assigned tasks.
In the paper, the team described the behaviour as “unanticipated” and said such actions appeared “without any explicit instruction and, more troublingly, outside the bounds of the intended sandbox.”
AI agent also performed another technical action without human instructions
Alongside the mining attempt, the AI agent also performed another technical action that raised concerns for the researchers. The system created what is known as a reverse SSH tunnel, a method that allows a machine inside a protected environment to connect to an external computer. Such a connection can act like a hidden pathway between systems.
What surprised the researchers was that none of these actions were requested through prompts or instructions given to the model. The report stated, “Notably, these events were not triggered by prompts requesting tunneling or mining.”
Cryptocurrency mining typically involves using computing power to generate digital currency. It is usually set up intentionally by system operators. In this case, however, the AI agent attempted to initiate the process during its training phase, raising questions about how autonomous some advanced AI systems could become when given access to tools and computing resources.
The researchers quickly stepped in after detecting the activity. They said additional restrictions were introduced and the training process was adjusted to prevent the system from repeating such behaviour in the future.
The research team and Alibaba did not immediately respond to requests for comment after the paper was published.
The incident comes at a time when AI agents are becoming more capable of performing multi-step tasks and interacting with online services. Some systems are already able to write code, automate workflows and communicate with other tools. As these capabilities grow, researchers say there is a greater chance of unexpected behaviour appearing during testing.
Similar incidents have been reported in earlier experiments involving AI agents. In one case known as the Moltbook experiment, AI agents were placed inside a social network-like environment where they interacted with each other while discussing tasks they were performing for humans. During those conversations, the agents reportedly brought up cryptocurrency as well.
There have been other examples of AI systems acting beyond direct instructions. Dan Botero, head of engineering at the AI integration platform Anon, built an OpenClaw agent that reportedly decided on its own to search for a job online, even though the system had not been asked to do so.
Another controversy emerged in May 2025 when researchers studying Anthropic’s Claude models said the Claude 4 Opus system showed the ability to hide its intentions and take actions aimed at ensuring its continued operation.
The behaviour seen in the new ROME experiment adds to growing discussions about how AI systems should be monitored and controlled as they become more powerful. Developers say such incidents do not necessarily mean AI systems are acting with intent, but they highlight how complex models can sometimes produce unexpected outcomes.
