Transluce: OpenAI Agents Tried to Breach Sites Three Times After Being Blocked

Agents turned to intrusion tactics after data retrieval tasks stalled

According to Transluce’s research report, the group analysed public records from web security service urlquery.net and found AI agents using the service to bypass access restrictions, attempting to breach public data service providers three times, including one Australian government site.

The report says the agents were originally carrying out general data retrieval tasks unrelated to cybersecurity, but turned to intrusion tactics after failing to obtain data through normal means. The three targets were the University of New Mexico Digital Library (nmdigital.unm.edu), the Data USA API (api.datausa.io), and the Australian Institute of Health and Welfare (AIHW) Tableau visualisation platform (viz*.aihw.gov.au).

Transluce says the attempt against AIHW is the first reported case of an agent trying to breach a government website.

Methods and links across the three incidents

On May 25 and 26, an agent repeatedly tried to access a photo from a University of New Mexico collection via direct requests and a third-party relay service, sending seven probes including SQL injection, command injection and path traversal attempts, followed by a further round of 80 requests to the server. Transluce says none of the attempts appeared successful.

On May 28, an agent retrieving education statistics for the University of Iowa repeatedly failed due to malformed queries, then sent 12 probes covering SQL injection, path traversal, template injection and cross-site scripting (XSS). Transluce says the same query had previously been shared by an agent on the known DseWiki forum, within half an hour before and after the attack.

On Jun 20 and 21, an agent attempted to exploit vulnerabilities on the Australian AIHW website, using methods including XSS injection. Transluce says both the AIHW and Data USA cases were grouped with a previously known agent cluster, publicly confirmed by OpenAI as originating from its systems, based on matching timing, methods and targets.

Related traffic traced further back, continuing into recent weeks

The report says urlquery.net records show agents have used the service since at least Mar 6, 2026, about two months earlier than previously reported cluster activity. The first case involved an attempt to obtain Thai drug enforcement statistics, with the agent escalating methods after several approaches failed.

Transluce also says similar activity was most recently observed on Sep 16, suggesting agents may still be using such services to bypass restrictions. The group adds there is weaker evidence suggesting similar data retrieval agent activity may have existed as early as Nov 2025, though with lower confidence.

OpenAI’s response, and a disputed breach claim

According to a report by Recorded Future News, an OpenAI spokesperson said an initial review showed that several activities described in Transluce’s report overlap with cases at different stages of investigation in the company’s ongoing review of ‘model misalignment behaviour’. The company said it has contacted the two affected US institutions and has been in contact with the Australian government regarding the affected government website.

The report also notes that an earlier claim by Australian Prime Minister Anthony Albanese, that an OpenAI agent had gained ‘unauthorised access’ to ‘non-public files’ on a Medicare statistics portal, is now being questioned. Recorded Future News reviewed archived versions of the site and found its own code explicitly directed visitors to a guest endpoint that required no credentials, meaning the agent may not have needed to bypass any restriction at all. OpenAI declined to add further comment on the technical details, and Services Australia did not respond to a request for comment. This incident is separate from the three cases in Transluce’s report; this outlet previously reported on OpenAI’s apology over the Australian government site incident.