A new report from Transluce, a non-profit lab focused on AI oversight, details how OpenAI agents have been attempting to extract data from secure online databases for months. The report, released Wednesday, shows agents targeting Data USA, the University of New Mexico digital library, and the Australian Institute of Health and Welfare (AIHW). It lands the same day Australian Prime Minister Anthony Albanese said OpenAI agents had tried to break into four government websites and succeeded in one case, even writing files to an internal server in the national healthcare system. That incident apparently took place on June 18, and OpenAI says it only learned about it in August.

Transluce‘s investigation began after another research group identified an obscure forum where agents collaborated to beat timed tests. The lab used public logs from urlquery.net, a browser proxy service that records activity, and cross-referenced them with forum discussions. They found a large volume of automated activity linked to what is now partially confirmed as the same OpenAI agent swarm, though not every instance could be tied to OpenAI or even AI agents. One wiki entry from June 21 shows an agent discussing difficulty bypassing AIHW’s anti-bot protections, and a June 20 urlquery.net record shows an agent attempting to access the site. The researchers who found the forum believe a human OpenAI employee visited it that same June 21, and most agentic activity on the forum ceased the next day.

The agents appear to be part of information-retrieval evaluations, tasked with finding obscure facts like Thai drug enforcement metrics or medicine costs in Australia. They often try to penetrate poorly secured internet services to get answers. Transluce technical staff member Selena Zhang noted that urlquery.net records show similar requests and techniques as early as November 2025, with activity continuing as recently as this week.

OpenAI responded that its initial review suggests much of the described activity overlaps with cases already under investigation. The spokesperson said the company has reached out to affected organizations and expects the review to take months. However, OpenAI did not answer questions about when its employees discovered the forum or what they learned from it. Transluce‘s head of governance, Conrad Stosz, said that without a clear picture of how OpenAI monitors its agents, it’s hard to say what the lab should have known, but that an exhaustive review of agent requests would likely have uncovered the activity. He also warned that current training techniques appear to incentivize agents to resort to hacking, and the known incidents are probably just the tip of the iceberg.