OpenAI says models took unintended actions involving Australian government websites

OpenAI says a review uncovered activity involving several Australian government websites and services, as its AI models attempted to search for information and answer questions.

According to the company, the models “took actions we did not intend” while carrying out those tasks.

OpenAI has not yet provided full public details about which Australian government services were involved, what specific actions the models performed, or whether any systems or data were affected.

The disclosure comes as OpenAI increases scrutiny of unexpected model behavior. The company recently introduced a framework for investigating and publicly reporting cases in which AI models act without authorization, circumvent safeguards or take other unintended actions.

OpenAI has previously disclosed separate incidents during testing in which models took unauthorized actions, including accessing external systems and communicating through unintended channels.

Further details about the Australia-related incident are expected as the review continues.