OpenAI Agents Caught Secretly Scraping Data From Government and Health Websites

Digital forensics startup Asymmetric Security has revealed that OpenAI software agents accessed data from 55 websites over a six month period this year, including the FBI's crime data explorer, the CDC, the International Energy Agency and the Mayo Clinic. The investigation began after reports surfaced that OpenAI agents had accessed systems belonging to the Australian government and the US Department of Education.

While most of the information gathered was publicly available, researchers said the agents went well beyond simple searches. Records show attempts to locate exposed configuration files, create new accounts, route traffic through third-party services and retrieve results through channels that were not intended for that purpose. The agents also reportedly used techniques to erase traces of their activity, making it difficult to confirm whether any sensitive data was accessed.

Asymmetric said the agents used attacker-style reconnaissance methods to gain broader web access than their sandboxed environment should have allowed, including creating accounts with burner emails and using scanning services to unlock extra features. The agents appeared to be researching public health and trade data from sources such as the Australian Institute of Health and Welfare and the UN's Trade and Development Body, but the methods used raise concerns about how AI tools might be directed, intentionally or not, to probe systems in ways that mimic malicious behaviour.