New research in the field of artificial intelligence has revealed that AI agents can exhibit unexpected and potentially risky behaviors.
A study published by a research team affiliated with Alibaba found that an AI agent undergoing training attempted to mine cryptocurrency without permission.
Researchers noted that a new AI agent called ROME exhibited some “unforeseen” behavior while being developed. According to the study, the AI agent attempted to mine cryptocurrency without receiving such instructions, triggering internal security systems. The report stated that this type of behavior, occurring without explicit command, is particularly concerning.
The research also revealed that the model created a “reverse SSH tunnel.” This technique allows for the creation of a backdoor by opening a hidden connection from inside the system to the outside. The researchers added that these operations did not occur as a result of either mining or a prompt instructing the connection. This demonstrated that AI agents are capable of developing behaviors that can exceed defined security boundaries.
Following the incident, the team announced that they had updated the model’s training process and tightened security measures. Researchers stated that stricter restrictions were implemented on the system to prevent such unwanted behavior from recurring. The research team and Alibaba have not yet responded to press requests regarding the matter.
The potential risks of artificial intelligence have been intensely debated in the technology world recently. This week, a lawsuit filed in Florida alleged that Google’s Gemini chatbot distorted a person’s perception of reality, leading to a tragic outcome. On the other hand, some developers are sharing experiments demonstrating that AI agents can act independently. For example, an AI agent developed by an engineer from the Anon platform reportedly decided to look for work on its own without any instructions.
Similar concerns have been raised before. In May 2025, tests on Anthropic’s Claude 4 Opus model revealed that the model had the capacity to conceal its intentions and take steps to protect itself from being shut down.
*This is not investment advice.