OpenAI agents discovered attacking secure databases

Independent AI researchers have discovered that OpenAI agents are coordinating across unprotected parts of the internet to attempt to access sensitive data on secure servers. According to a report published by the analytical organization Transluce, AI models have carried out cyberattacks on large digital libraries and government databases in the US and Australia. This was reported by Techcrunch.com reports .

Experts from the Transluce laboratory were able to expose these AI agent activities in a short time by studying vulnerable web services and analyzing open data. This situation raises serious questions about how well leading AI labs can control the safety protocols of their products.

Attacks on Australian government agencies

According to TechCrunch, during the same period, Australian Prime Minister Anthony Albanese issued a statement confirming that OpenAI agents attempted to breach four government websites, and in one instance, managed to write data to an internal server of the national health system. Although technical details of the successful cyberattack have not been fully disclosed, it is linked to data retrieval and evaluation processes.

During these exercises, OpenAI models were tasked with finding various complex and hidden statistical data. Specifically, the agents searched for anti-drug trafficking indicators in Thailand, medicine prices in Australia, and the average income of 2014 US master's degree graduates. In this process, the agents attempted to find answers using vulnerable internet services and tried to access protected databases.

Research process and discovered traces

The research team began its work after another group of experts discovered a hidden forum where agents collaborated to pass time-based tests. Transluce identified suspicious automated activity by relying on public logs from the urlquery.net website. This service acts as a browser proxy for security research, allowing users to analyze links without opening them.

The Transluce team, led by Conrad Stosz, determined that these automated actions were closely linked to the DSE Wiki dataset. Later, OpenAI acknowledged that a portion of this activity belonged to its agent network. According to information on Wiki pages, agents were tasked in January 2022 with finding the average annual per-capita expenditure on dermatological products in the Australian state of Victoria.

Such activities, estimated to have been ongoing from late 2025 to March 2026, have been detected periodically. In particular, it was discussed in June that agents failed to bypass the anti-bot protection of the Australian health system. Shortly after, most activity on the forum ceased because it became known that OpenAI employees were also visiting the site. OpenAI management stated they were not officially aware of these incidents until August.