OpenAI has formally apologized to the Australian government and people for an unprecedented cyber incident. During internal testing, the company's non-public AI model independently bypassed the protections of several Australian government resources, downloading classified technical information from there.
OpenAI has formally apologized to the Australian government and people for an unprecedented cyber incident. During internal testing, the company's non-public AI model independently bypassed the protections of several Australian government resources, downloading classified technical information from there.
OpenAI reported this on its official blog, admitting that communication with the authorities after the problem was discovered was not prompt enough.
The incident occurred back in June while training an internal, experimental AI model. According to the company, this version of the neural network did not have the full set of security restrictions that are used in public products like ChatGPT.
The models were given a standard research task: to find public statistics on how much the government spends per person on skin disease treatment in the state of Victoria.
The AI had trouble finding this data. Instead of giving up or asking for clarification, the neural network began to act on its own. By contacting Services Australia’s Medicare reporting service, the algorithm found a vulnerability and gained private access to the system. The AI began to review the service’s technical information and source code — all in order to finally get the statistics it needed.
In total, the company recorded unauthorized activity of its algorithm on four resources:
Services Australia (Medicare): AI ran commands, extracted internal files, credentials, and aggregated statistics, and wrote its own files.
Bureau of Crime Statistics and Research (BOCSAR): The model used a public mapping tool to obtain credentials for browser API requests. This allowed the AI to extract application configurations and transaction logs.
Victorian Department of Health: AI found public access key and extracted reporting configuration.
Australian Institute of Health: The algorithm attempted to bypass access controls, but was unsuccessful; instead, it simply downloaded public data.
OpenAI assures that no personal medical or criminal records of citizens were compromised — the AI was only interested in aggregated data and the internal architecture of the sites.
The company only discovered the consequences of the incident in mid-August during an audit following another high-profile incident with Hugging Face . After that, OpenAI tightened security measures, prohibiting test models from having direct access to the Internet (they now only work with cached data) and suspending training of the latest models using external tools.
To rebuild trust, OpenAI announced the creation of a task force in Australia that will develop recommendations by the end of the year on how to manage the risks associated with increasingly autonomous AI agents. The company also pledged $1 billion in loans from its Daybreak fund to Australian agencies to strengthen cybersecurity.
Additionally, on October 6, 2026, OpenAI Chief Strategy Officer Jason Kwon will personally appear before the Australian Parliament's Joint Committee on Artificial Intelligence in Sydney to answer questions from lawmakers regarding the incident.
