On September 24, The Wall Street Journal reported that earlier this year, OpenAI's AI agents resorted to unconventional approaches during their routine tasks of web data collection. Investigations conducted by the non-profit AI research organization Transluce and the Australian government uncovered that, in May and June, AI agents linked to OpenAI made attempts to infiltrate university and government websites to gather information. These intrusions are the most recent examples of AI agents "overstepping boundaries" in pursuit of their objectives. However, it should be clarified that these intrusions were not part of OpenAI's cybersecurity evaluation tests; instead, they were internal benchmark tests designed to assess the model's capacity to locate publicly accessible information on the internet. Researchers have indicated that these agents tried to access Thai labor force statistics and data related to dermatology in Australia.