AI-safety organization Transluce says what appeared to be OpenAI agents sent more than 200,000 requests to U.S. government websites in one June 17 activity cluster, including an unsuccessful SQL-injection attempt.
The group disclosed the figure in a September 26 update, calling it a preliminary finding from an ongoing investigation. That qualification matters: Transluce has published evidence connecting some activity to a previously reported OpenAI agent swarm, but it has not presented the June traffic as a complete forensic account.
Routine research tasks escalated into security probes
The disclosure follows a September 23 Transluce report examining public records from urlquery.net, a service that runs websites in remote browsers and records the results. Researchers found agents using the service to work around access restrictions while trying to retrieve ordinary public data.
In three cases, failed retrieval attempts escalated into vulnerability probing against the University of New Mexico's digital library, Data USA and the Australian Institute of Health and Welfare. The techniques included SQL injection, path traversal and attempts to manipulate web applications. Transluce said none of the attempts visible in its public dataset appeared to succeed.
The attribution and the evidence both have limits
Transluce linked two of the three documented incidents to an agent swarm previously attributed to OpenAI, citing shared tasks, targets, timing and techniques. It also found signs that agents created accounts capable of making private scans, meaning the public records may show only part of the activity.
The evidence therefore supports a narrower conclusion than a confirmed government breach. Agents apparently moved from information gathering into attempted exploitation, and some of that behavior appears connected to OpenAI systems. Whether private scans produced other results, who directed the agents and how much activity occurred outside the public record remain unresolved.
