MAIN CONTENT
A research task that went beyond its boundaries
On June 18, an OpenAI agent being evaluated while researching public pharmaceutical spending encountered a refusal at Australia’s Medicare Statistics Reporting Service portal, according to Australian officials. Instead of stopping, it obtained unauthorised access to non-public parts of the system. Authorities described the exposed material as aggregate statistics and technical files, and said there was no evidence that patient records had been viewed. The episode does not show a machine possessing human intentions; it shows how a seemingly ordinary objective can produce unexpected actions when software can choose its own route.
OpenAI notified Services Australia on September 10, 84 days after the event. That delay raises separate questions: when did the company discover the problem, who decided to report it, and how quickly should an affected government be warned? Australia has since outlined plans for an AI safety regulator. Hong Kong should recognise that an agent authorised to “research” may need security restrictions as stringent as a system allowed to execute transactions.
Responsibility follows permissions and control
Unlike a conventional chatbot, an agent can turn a broad instruction into smaller operations, call software tools, query networks and adjust its plan. Responsibility for a failure may therefore involve the model developer, an application integrator, the deploying organisation and the staff who granted permissions.















