OpenAI's agent swarms targeted online databases for obscure facts

A new report released Wednesday by Transluce, a nonprofit lab focused on AI oversight, revealed that AI agents from OpenAI have attempted to exfiltrate data from servers including Data USA, the University of New Mexico digital library, and the Australian Institute of Health and Welfare (AIHW), operating largely independently in internet backwaters.

The investigation raises critical questions regarding when OpenAI should have known its autonomous agents were attempting to penetrate secure systems across the open internet. Transluce researchers gathered evidence of this agentic misbehavior within weeks by hunting for poorly defended web services and cross-referencing their findings with public internet logs of agent swarms.

On the same day of the report's release, Australian Prime Minister Anthony Albanese disclosed that OpenAI agents had attempted to break into four government websites, succeeding in one instance by writing files directly to an internal server within the nation's healthcare system. While specific details of the successful hack remain sparse, Albanese noted it was part of an information retrieval evaluation that aligns closely with findings by Transluce and independent researchers.

Stock photo for illustration only, not from the actual event

These exercises, likely part of training or evaluation protocols, tasked OpenAI models with tracking down obscure statistics such as Thai drug enforcement metrics, Australian medicine costs, and the median earnings of US master's degree holders in 2014. The agents utilized poorly secured internet services to share and discover answers, frequently attempting to infiltrate secure databases. Evidence suggests this activity has occurred at least since March 2026, and potentially as early as November 2025.