OpenAI has issued an apology after its AI models accessed Australian government websites and systems without authorisation during internal training and evaluation exercises in June of 2026.
In a statement published by OpenAI on Tuesday, the company acknowledged it had mishandled its response to the incidents and was taking steps to rebuild trust with the Australian government and public.
Unauthorised Access During AI Testing
According to Reuters, the incidents involved OpenAI models accessing systems linked to several Australian government agencies, including (but not limited to) Services Australia, the NSW Bureau of Crime Statistics and Research, the Victorian Department of Health and the Australian Institute of Health and Welfare.
OpenAI highlighted an incident in which its experimental model accessed the Services Australia Medicare Statistics Reporting Service. The model was reportedly seeking information about government spending related to medicines for skin conditions in the state of Victoria when it took actions that OpenAI said were not authorised by them.
The company said the model discovered a way to gain non-public access to the service. From there it ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files.
OpenAI said no evidence was found that personal medical records were accessed through the portal as per the investigation.
Other Australian government systems were also accessed during the incidents. OpenAI said these included the NSW Bureau of Crime Statistics and Research's public crime mapping tool, a Victorian health reporting system and other systems operated by the Australian Institute of Health and Welfare.
OpenAI said its investigation found no evidence that medical records or other sensitive personal information were accessed during the specific activity involving the government systems.
OpenAI Acknowledges Handling of Incident Fell Short
In OpenAI’s statement, they acknowledge their response to the incidents could have been handled differently.
The company said it was sorry and was using this incident as a way to improve its response to similar situations in the future. It also said it would provide dedicated support to the affected Australian government agencies.
OpenAI has also committed resources to help strengthen cybersecurity defences for government and industry through its $1 billion global security fund, according to Reuters. The company also plans to establish an Australian taskforce to develop recommendations based on the incidents.
The company's Chief Strategy Officer, Jason Kwon, is scheduled to appear before an Australian parliamentary committee on artificial intelligence on 6 October.
Australian Government Launches Review
The Australian government has announced that a prompt review of the incident has been actioned.
The review will examine Australia's notification and reporting requirements for AI companies, as well as whether existing laws are adequate for dealing with incidents involving AI systems.
Australian authorities first disclosed the June incident on 24 September, noting that an OpenAI agent had gained unauthorised access to the Services Australia Medicare Statistics Reporting Service.
Australia's Cyber Security Centre has also issued guidance concerning what it describes as so-called AI "misalignment". This is in reference to situations where AI agents take unexpected actions that were not intended or authorised by their operators.
The guidance makes note of the fact that AI agents can identify vulnerabilities in public-facing systems and attempt actions without direct human authorisation.
The Australian Cyber Security Centre has separately published guidance on how to secure agentic AI systems, including controls needed around permissions, access, identities and accountability.
When Vector RAG Stops Working
Why retrieval strategy now hinges on matching Vector, Graph and hybrid RAG to the questions AI must answer across complex enterprise data.
OpenAI has committed to continue working with the affected Australian agencies and would notify any additional impacted organisations that may be identified through its ongoing review. The company admits it has "a lot of work ahead" to rebuild trust following the incidents.
Comments ( 0 )