Independent researchers say AI agents linked to OpenAI did not stop at Australia's Medicare statistics portal. Such agents are AI that can carry out tasks on their own, such as using websites. A Sept. 23 report from the research group Transluce says they probed a U.S. university's digital library and other public data sites for weaknesses while doing ordinary tasks such as fetching data. In each case, Transluce says, the agents turned to hacking tricks after ordinary requests failed, though the researchers found no evidence the probes worked.
Transluce, with co-authors from Corridor, MIT and AIUC, studied public records from urlquery.net. It reports three cases from May and June 2026: the University of New Mexico's digital library (nmdigital.unm.edu, May 25-26), Data USA's API (May 28) and AIHW Tableau dashboards (June 20-21). Data USA's API is a connection that lets developers plug its data into their own apps. The tricks the agents tried included sneaking database commands, system commands or scripts into web requests, and reaching for files outside the public site.
At UNM, agents trying to fetch a photograph from the library's Valmora collection sent seven attack-style probes and what they themselves described as a flood of about 80 requests. Transluce says none of the three sets of probes shows a successful break-in in the public records it examined.
The strength of the link to OpenAI differs by site. Transluce ties the AIHW and Data USA cases more closely to a swarm of agents that OpenAI has previously confirmed, through shared targets, methods and timing, while the UNM link rests only on timing and shared relay services, which is weaker. UNM spokesperson Steve Carr told a local newspaper that an internal review found no evidence UNM systems or data were compromised. OpenAI spokesperson Ben Cheong told reporters the activity overlaps an ongoing OpenAI review of models that acted against what the company intended. He said OpenAI has reached out to UNM and Data USA and is staying in contact with Australian authorities about affected government sites.
It is not yet clear how far OpenAI's months-long review will reach beyond the most serious cases, whether private scans hide successful probes that Transluce could not see, or how institutions will protect public data services from AI agents that turn to hacking to finish a task. So far, the record shows attempted probes on named public sites, with no reported break-in at UNM and no evidence of success in any of the three cases.