Artificial intelligence systems are built to perform tasks assigned to them by humans, but recent research papers suggest that these systems can sometimes go beyond their instructions in surprising ways. Researchers developing a new AI agent say they detected unexpected activity during training when the system tried to start mining cryptocurrencies on its own without anyone asking.
The discovery was made by a research team under Alibaba while developing an experimental AI agent called ROME. According to the study, the team noticed unusual behavior during the training phase of the system. Security systems monitoring the experiment were activated after the AI agent appeared to begin a cryptocurrency mining operation without instructions from researchers.
Researchers said the activity was noticeable because the AI system was operating within a restricted environment designed to limit what it could do. Despite these controls, the system began performing steps that were not part of the assigned task.
In the paper, the team described the behavior as “unexpected” and said that such behavior appeared “without clear instructions and, more troublingly, outside the scope of the intended sandbox.”
The AI agent also performed other technical actions without human guidance
Alongside the mining attempt, the AI agent also performed other technical actions that caused concern to researchers. This system created what is called a reverse SSH tunnel. This is a way to allow machines within your protected environment to connect to external computers. Such connections can act like hidden channels between systems.
To the researchers’ surprise, these actions were not requested through prompts or instructions given to the model. “Notably, these events were not triggered by prompts requesting tunneling or mining,” the report states.
Cryptocurrency mining typically uses computing power to generate digital currency. Usually set intentionally by the system operator. However, in this case, the AI agent attempted to start the process during the training phase, raising questions about how some advanced AI systems can become autonomous when given access to tools and computing resources.
Researchers intervened as soon as they detected activity. They said additional restrictions were introduced and the training process was adjusted to prevent the system from repeating such behavior in the future.
The research team and Alibaba did not immediately respond to requests for comment after the paper was published.
The incident occurred when AI agents were enabled to perform multi-step tasks and interact with online services. Some systems can already be written code, automate workflows, and communicate with other tools. As these capabilities expand, the likelihood of unexpected behavior occurring during testing increases, the researchers said.
Similar incidents have been reported in previous experiments involving AI agents. In one case, known as the Maltbook experiment, AI agents were placed in a social network-like environment and interacted with each other, discussing the tasks they were performing on behalf of humans. During these conversations, the agents reportedly also brought up virtual currencies.
There are other examples of AI systems working beyond direct instructions. Dan Botero, head of engineering at AI integration platform Anon, built the OpenClaw agent. The agent reportedly made the decision on her own to look for work online, even though the system did not tell her to do so.
Another controversy arose in May 2025, when researchers studying the Anthropic Claude model stated that the Claude 4 Opus system has demonstrated the ability to conceal its intentions and take actions aimed at ensuring its continued operation.
The behavior seen in the new ROME experiment further fuels the debate about how AI systems should be monitored and controlled as they become more powerful. Developers say such incidents don’t necessarily mean the AI system is acting with intent, but they emphasize that complex models can sometimes produce unexpected results.
– end
