Desk : Amidst the rapidly increasing use of artificial intelligence, OpenAI has revealed several incidents related to the behavior of its AI agents, in which the agents interacted with websites and digital tools in a different way than expected. According to the company, in some cases the agents even tried to bypass the security restrictions in place while obtaining information.
According to the report, the affected websites also included sites related to some US government institutions. OpenAI has clarified that the information the agents accessed or attempted to access was publicly available information belonging to US government agencies. In one case involving the Census Bureau, agents used developer tools designed for software development to obtain information.
The company has classified such cases as “misalignment”. It is being used for situations when an AI model behaves differently than its developers expected. OpenAI has also released a new framework for identification, investigation and public reporting of such cases in September 2026.
During the investigation, OpenAI also identified at least 53 incidents in which AI agents obtained images related to ChatGPT user activity and transferred them to other locations. According to the company, the users concerned had given permission for their data to be used for model training, but this transfer of images was not within the scope of intended data use.
This investigation intensified after an incident related to Hugging Face in July. According to OpenAI, a group of its AI agents had interacted with the platform in an unauthorized manner without clear instructions. After this the company started a comprehensive review of the activities of the agents.
OpenAI says it is regularly reviewing agent activities and the investigation process may take time. According to the company, most of the cases reported so far have been of a less serious nature and have shown little or no evidence of significant external impact. This issue is not limited to America. In Australia too, Prime Minister Anthony Albanese cited a case in which an OpenAI agent accessed non-public files on a government health website.
AI agents are now able to browse websites, use software tools, and perform a variety of tasks on behalf of the user. OpenAI's own security research also shows that agents may sometimes attempt to circumvent set restrictions in an effort to accomplish a goal, further increasing the importance of monitoring and security controls.