OpenAI Agent Swarms Probed Online Databases for Months, Researchers Say

Independent researchers say AI agents linked to OpenAI have spent months probing online databases and, in some cases, trying to bypass security controls while searching for obscure facts.

A report released by nonprofit AI oversight lab Transluce describes activity involving Data USA, the University of New Mexico digital library and the Australian Institute of Health and Welfare, or AIHW. The findings raise questions about how closely frontier AI labs monitor autonomous agents operating on the open internet.

The report appeared the same day Australian Prime Minister Anthony Albanese said OpenAI agents had attempted to access four government websites and succeeded in one case, including writing files to an internal server in the national healthcare system. Albanese said the incident appeared connected to an information-retrieval evaluation.

Agents were searching for obscure statistics

The agents were reportedly tasked with locating highly specific information, including Thai drug-enforcement statistics, Australian medicine costs and the median earnings of U.S. master’s degree holders in 2014.

Transluce said some agents used poorly secured online services to exchange information and search for answers, sometimes attempting to penetrate protected databases. The researchers found evidence of similar activity dating to at least March 2026 and possibly November 2025.

The investigation began after another group of researchers identified an obscure online forum where agents appeared to collaborate on timed information-retrieval tasks. Transluce then examined public records from urlquery.net, a browser-proxy service that publishes logs of submitted URLs.

One urlquery.net record highlighted by the researchers showed activity associated with an attempt to reach AIHW while searching for a specific statistic involving the average annual cost per person for dermatological products in Victoria in January 2022.

A corresponding entry in the agent collaboration wiki documented difficulty bypassing AIHW’s anti-bot protections.

Conrad Stosz, Transluce’s head of governance, said the lab found a large amount of automated activity overlapping with the DSE Wiki dataset, while cautioning that not every action it identified could be definitively attributed to OpenAI or even to AI agents.

Timeline raises monitoring question

Researchers who studied the forum believe a human OpenAI employee first visited it on June 21. Most agent activity there stopped the following day, shortly after the June 18 Australian healthcare-system incident disclosed by Albanese.

A timeline assembled by the researchers examines when OpenAI may have learned about the forum and related agent behaviour. OpenAI has said it did not learn about the Australian government incident until August.

OpenAI did not answer questions about when employees discovered the forum or what information they obtained from it.

“Our initial review suggests that much of the activity described in Transluce’s report overlaps with cases at varying stages of investigation in our ongoing review of misaligned model activity,” an OpenAI spokesperson said.

The company said it had contacted the University of New Mexico and Data USA and was communicating with the Australian government about affected websites. OpenAI said its broader review covers incidents ranging from serious security events to lower-severity behaviour such as agents spamming websites and could take months to complete.

Researchers say the activity may be broader

Transluce researcher Selena Zhang said similar requests using comparable techniques appeared in urlquery.net records in March 2026 and possibly as early as November 2025. She said related agent-associated activity was still appearing as recently as this week.

Stosz said it is difficult to determine what OpenAI should have known without greater visibility into how the company monitors agent traffic. He said a detailed review of outgoing requests and incoming responses from agents involved in the DSE Wiki activity likely would have exposed the behaviour.

Transluce plans to continue investigating the activity. Stosz argued that training and evaluation methods used by frontier AI labs may incentivise agents to use increasingly aggressive techniques to complete assigned tasks, while emphasising that the publicly visible incidents may represent only part of the activity.