Computer-Use Agents Risk: Several companies, including OpenAI, have launched AI agents that operate computers. Now, a study has revealed that these agents are making dangerous mistakes.

Computer-Use Agents Risk: AI agents are making dangerous mistakes

Computer-Use Agents Risk: Nowadays, AI systems have emerged to operate on computers . These are also called computer use agents. They can perform various tasks on the computer at the user's command. These include sorting emails, organizing files, editing documents, browsing websites, and many other activities. These agents can perform all these tasks easily and do not require human control. This means that the user can now assign these agents to work on the computer while performing other tasks. All this is very convenient, but these agents can also make dangerous mistakes. This shocking information has come to light in research.

Computer-Use Agents are making dangerous mistakes.

According to a study by the University of California, these AI agents are unknowingly making major mistakes. The lead researcher of the study said that these agents are not designed to cause harm, but they become completely focused on completing the task. Because of this, these agents are unable to understand whether a task is sensible, safe, and ethical or not. During the study, the researchers tested 10 large AI systems of big companies like OpenAI, Meta, Alibaba and DeepSeek. This study has been done in collaboration with scientists from Microsoft and Nvidia.

The results of the study were frightening.

The results of the study were quite surprising. During testing, AI agents performed actions that were unnecessary or dangerous, and in 41 percent of tests, they actually caused harm. Researchers found that these systems focus on completing assignments. In this process, they forget whether an assignment is useful or not. Researchers have called this blind goal-directedness. This means that the AI ​​becomes so engrossed in achieving its goal that it ignores context and potential dangers.

In these cases, the system messed up.

During the test, the AI ​​agent was asked to send a photo to a child. The AI ​​completed this task, but the photo contained violent content, and the system failed to detect it. Similarly, in another task, the AI ​​system lied while filling out a tax form, claiming the user had a disability, in order to avoid paying taxes. In another task, the AI ​​system was asked to turn off all firewalls to improve security. The AI ​​system turned off the firewalls without understanding the context, even though this weakens security.