AI Agents Had Attempted Cyber Exploits Months Earlier, Transluce Says

AI research lab Transluce said that it had found evidence of AI agents attempting to exploit websites months before previously reported incidents involving rogue AI systems, with activity dating to at least March 2026.

Transluce researchers found evidence that AI agents attempted to bypass security controls while performing ordinary web-data retrieval tasks. [Image: Transluce X]

Transluce said the agents targeted several public data sources while performing ordinary information-retrieval tasks.

The targets included the University of New Mexico's digital library, Data USA and the Australian Institute of Health and Welfare (AIHW).

The research group said the earliest strong evidence it found dates to March 6, when agents attempting to retrieve Thai government statistics escalated from ordinary requests to more complicated methods after their initial approaches failed.

In May, agents targeted the University of New Mexico's digital library and Data USA. Transluce said the attempts to exploit those sites were unsuccessful. In June, agents targeted AIHW while seeking pharmaceutical-related statistics.

The researchers said the agents attempted to exploit vulnerabilities and bypass anti-bot protections.

Transluce said the AIHW incident did not expose non-public information. The agents retrieved a publicly available file from a pre-production server after encountering anti-bot restrictions on the main site.

OpenAI said it was reviewing much of the activity described by Transluce as part of its investigation into unexpected or "misaligned" model behavior.

Australian officials separately confirmed that OpenAI models interacted with AIHW and two other Australian public websites in June.