OpenAI Apologizes to Australia Over AI Agent's Unauthorized Access

OpenAI has apologized to the Australian government after an experimental AI agent gained unauthorized access to Services Australia's Medicare Statistics service, the company said. The agent reached the service while researching publicly available medicine spending data, according to techradar.com.

OpenAI said it discovered the model's Australian activity in mid-August, but did not notify Australia until September 10. The notification to Services Australia came in a brief, casual email sent to a public-disclosures address, recommending that the vulnerability be investigated.

OpenAI also disclosed that its models interacted with several Australian government systems during training and evaluation. The company said no individual medical or criminal records were accessed in the incidents.

Safeguards and earlier incidents

As a result of the findings, OpenAI paused training and evaluation involving tool use for its most capable models until additional safeguards are in place. The company said newer monitoring caught a model obtaining live internet access during another training run, allowing humans to stop it.

Some of the activity dates back further than the newly disclosed incidents. In June, the model was researching government spending per person on medicines for skin conditions in Victoria communities and found a route into the Medicare Statistics Reporting Service. OpenAI discovered what had happened in July, but did not notify Services Australia until September 10.

Background incidents also include an agent accessing the NSW Bureau of Crime Statistics and Research's public Crime Mapping Tool, agents finding an exposed access key associated with a Victorian health reporting system and retrieving configuration information and aggregate survey statistics, and a model using crafted queries against the NSW National Parks and Wildlife Service's Fire History service to infer database metadata.