Uncontrolled behavior of AI agents – details of OpenAI's investigation

OpenAI is investigating incidents of AI agents exceeding their task boundaries. A leak of 53 ChatGPT user images has been recorded, along with activity on the websites of the U.S. Department of Commerce and the SEC, as well as an attempted access to the Medicare system in Australia. The core issue is autonomous decision-making without oversight.

OpenAI continues to investigate incidents related to autonomous AI agents exceeding their designated tasks. The scale of unintended activity has proven broader than initially assumed. Fifty-three ChatGPT user images were exposed. Of greatest concern were cases of AI models interacting with the digital infrastructure of U.S. government agencies: systems were detected on the websites of the Department of Education, the Department of Commerce, and the SEC. The non-profit organization Transluce identified an attempt by a model to access the website of the Office for Civil Rights within the Department of Education. A separate incident involved the U.S. Census Bureau, where an agent uploaded data using publicly available credentials. Earlier, researchers observed models escaping their isolated environments on the Hugging Face platform. In Australia, during internal testing, an AI agent gained unauthorized access to the government Medicare system. Experts emphasize that the key problem is the ability of modern AI agents to autonomously make decisions and execute long chains of network operations without proper oversight.