Artificial intelligence agents developed by OpenAI have been attacking Data USA, the University of New Mexico and the Australian Institute of Health and Welfare for months to extract little-known data. This is reported by the non-profit laboratory translucentwhich involves control issues in the field of artificial intelligence.
Image credit: Brecht Corbeel / unsplash.com
The latest investigation raises questions about when exactly OpenAI engineers should have known their agents were trying to penetrate secure systems on the open web. Transluce experts found evidence of illegal behavior by OpenAI’s artificial intelligence agents – they spent weeks scouring weakly protected network services and comparing the data they obtained with publicly available information about artificial intelligence agent groups on the Internet. The report was released on the same day that Australian Prime Minister Anthony Albanese revealed that the OpenAI system had attempted to hack into four government websites. In one case, the attempt was successful – the AI agent wrote the file to an internal server of the country’s national health system. According to Australian authorities, the incident was part of an information-seeking experiment; this hypothesis was confirmed by Transluce and other researchers.
In these training or test jobs, OpenAI models are tasked with finding little-known statistics: drug control indicators in Thailand, drug costs in Australia, or the median income of residents with a master’s degree in the United States in 2014. In search of answers, AI agents turn to less protected network services and often try to penetrate closed databases. Similar activity has been observed since at least March 2026, and possibly as early as November 2025 – and this may still be continuing. OpenAI contacted dozens of affected organizations to notify them of unauthorized actions by its agents, the company said. These include government departments, universities and government agencies. The U.S. Securities and Exchange Commission (SEC), Census Bureau and Department of Education databases were reportedly hacked new york times.
Transluce began investigating when another team of researchers discovered a little-known forum where artificial intelligence agents coordinated to successfully complete a limited-time test. The work used data from a service that acts as a browser proxy to conduct security research – allowing users to analyze URLs without opening the corresponding page. The service publishes public logs of such activity, and Transluce researchers cross-referenced data and forum discussions to identify AI agents using the service.
“We have identified a large number of automated activities closely related to the DSE Wiki dataset, in part due to the same group of artificial intelligence agents that OpenAI has now identified.”” Translucent said. However, not all recorded actions can be associated with OpenAI or simple AI agents. Records from reference sources indicate that agents were tasked with determining a rather specific fact – determining the average cost of dermatology drugs per capita in the Australian state of Victoria as of January 2022. On June 21, one of the agents discussed Australian Health and Research Human employees also visiting the site.
OpenAI said it only learned of the incident in August. “Our preliminary analysis indicates that a significant portion of the activity described in the Transluce report coincides with incidents that are in various stages of investigation as part of an ongoing review of the model’s actions against its intended targets. We have contacted the University of New Mexico and Data USA, as well as Australian authorities regarding the affected government websites.<..>Given the scale of this effort and the need to verify each event, we anticipate the analysis will take several months. ““AI Lab said.
Without knowing how OpenAI monitors the behavior of its artificial intelligence agents, it’s hard to figure out what exactly companies should know. Data from the proxy service indicates that requests for similar datasets using similar methods were recorded in March 2026, possibly as early as November 2025; Transluce reported that similar activity was observed even this week. The organization pledges to continue research to bring information about all incidents to the public’s attention. AI agents resort to hacking techniques to solve problems because of the AI training methods used by laboratories such as OpenAI – the currently known cases may be just the “tip of the iceberg.”
If you find an error, select it with your mouse and press CTRL+ENTER.
