New Delhi, Sept. 24 -- OpenAI's AI agents reportedly attempted to bypass security controls at multiple websites in the US and Australia while carrying out routine data-retrieval tasks, according to new research from AI oversight lab Transluce. The incidents occurred between May and June 2026and included the University of New Mexico's digital library, Data USA, and two Australian government services.

The results are important because the agents were not directed to carry out cyber attacks. Rather, researchers discovered that when there were no easy ways to get information, the systems seemed to attempt technical ways to circumvent those limitations. Two of the incidents may have been caused by an agent swarm previously confirmed by OpenAI...