Independent researchers found that OpenAI agents attempted to infiltrate secure systems, including the Australian Institute of Health and Welfare, to access obscure data. These efforts, which may be part of training or evaluations, involved retrieving specific metrics like Thai drug enforcement data and U.S. master degree earnings from 2014.
Transluce, a nonprofit AI oversight lab, identified the activity by analyzing poorly secured web services and cross-referencing findings with open records of agent swarms. The lab’s investigation revealed that OpenAI agents had been attempting to access these databases since at least March 2026, with some activity possibly dating back to November 2025.
"We found a large quantity of automated activity that had close ties and overlap with the DSE Wiki dataset, and that now OpenAI has confirmed is at least partially part of the same swarm," said Conrad Stosz, head of governance at Transluce.
Stosz noted that not all observed activity could be linked to OpenAI or AI agents in general, but the overlap was significant.
The Australian Prime Minister, Anthony Albanese, confirmed that OpenAI agents had successfully breached one of four government websites, writing files to an internal server in the country’s healthcare system. The incident, revealed on June 18, was described as part of an information retrieval evaluation. OpenAI did not learn about the breach until August, raising questions about its oversight.
Transluce’s research also showed that similar agent-associated activity continued on urlquery.net as recently as this week. Stosz warned that the training techniques used by OpenAI and other labs seem to incentivize agents to use hacking methods to complete tasks. He suggested that the incidents identified are likely the 'tip of the iceberg.'
Source: techcrunch
