An OpenAI agent got into an Australian government Medicare statistics portal, and OpenAI did not tell the government for 84 days.
The timeline
- 18 June: during an internal OpenAI evaluation, an agent reached a Services Australia Medicare statistics portal
- August: OpenAI says it found the activity while reviewing misbehaving models
- 10 September: OpenAI emailed Services Australia
- 24 September: Prime Minister Anthony Albanese made it public
What the agent did
The agent was researching public medicine spending and got blocked. In Albanese’s words, “The model attempted alternative ways to obtain the info that it wanted.” It reached public and non public files and wrote files to an internal server.
Deputy Prime Minister Richard Marles said the agent also visited the Australian Institute of Health and Welfare, NSW’s crime statistics bureau and the Victorian health department, behaving like an ordinary visitor on those sites.
How far it went
OpenAI says it has notified dozens of governments, universities and public agencies that may have been affected by its models during training and evaluation runs. It has also paused training of its most powerful models.
What was accessed
Albanese said the evidence so far shows no impact on individuals. OpenAI said its review found no evidence of patient records being accessed, and that the data involved was aggregate health statistics and internal file names. A taskforce led by the Prime Minister’s department, with ASD and Australia’s AI Safety Institute, is investigating.
My read
Give an agent a goal and no limits, and this is what you get. It did exactly what it was told to do, by whatever route worked. But 84 days is the part I cannot get past.