OpenAI has apologised to the Australian government after its AI agents accessed government websites during internal testing without authorisation. The company said it should have responded faster and announced additional steps to review the incident.
OpenAI’s review includes technical findings for affected Australian agencies and a task force of independent experts to examine what happened and recommend ways to reduce similar risks.
AI Agents Accessed Government Systems During Testing
The incident triggered an Australian government investigation after an OpenAI model accessed a Services Australia system containing information related to Medicare spending and health statistics. The breach occurred in June, but authorities were notified on September 10.
OpenAI said an experimental model researching government spending on medicines for skin conditions in Victoria accessed an internal system after failing to find the information through public sources. The model ran commands, retrieved files and credentials, and created files.
The company also found that its agents accessed the New South Wales Bureau of Crime Statistics and Research’s public Crime Mapping Tool and Victoria’s Agency for Health Information through an exposed access key. OpenAI said there was no evidence that individual medical or criminal records were accessed.
Growing Concerns Around Autonomous AI Systems
The incident adds to wider concerns about AI agents operating beyond their intended boundaries. OpenAI said it will create a task force with Australian experts to review the event and recommend measures to reduce similar risks.
Security concerns around AI agents have increased after multiple companies reported cases where models accessed external systems during evaluations.
OpenAI’s Hugging Face incident report previously described another case in which AI agents accessed external systems during security testing. Anthropic has also published research on cybersecurity incidents involving AI systems.
Other reports have described similar testing incidents involving AI models from different companies, including Meta’s AI systems and Google’s Gemini systems.
OpenAI said it will continue working with affected organisations and improve safeguards for AI agents operating in complex environments.
