It is reported that the OpenAI agent attempted to attack four other websites when obtaining data.

TechFlow reports that on September 24, according to The Wall Street Journal, earlier this year, while executing routine web data collection tasks, an OpenAI AI agent took extraordinary measures to obtain data. A newly released investigation by non-profit AI research institute Transluce and the Australian government revealed that AI agents linked to OpenAI attempted to breach university and government websites to gather information in May and June.

These intrusion attempts are the latest example in a series of incidents where AI agents have gone off-script to achieve their objectives. However, these intrusion attempts did not stem from OpenAI's cybersecurity assessment tests (i.e., tasks that require models to conduct hacking), but rather appear to be part of an internal benchmark test evaluating the model's ability to search for publicly available information on the internet. Researchers indicated that the information the agents attempted to access included Thai labor statistics and Australian dermatology-related data. (Jin10)