OpenAI's agent got into Australian health site

Robyn, anchorAn AI agent from OpenAI, the company behind ChatGPT, got into an Australian government health website without permission.

Robyn, anchorThat's according to Australia's prime minister.

Robyn, reporterAn AI agent is a program that pursues a goal by itself.

Robyn, reporterIn June, an internal OpenAI model was researching medicine spending.

Robyn, reporterThe AI agent found a way around those blocks. Didn't accept no for an answer, if you like.

Robyn, reporterThat's how Prime Minister Anthony Albanese puts it.

Robyn, reporterOpenAI says it ran commands and wrote files inside a statistics service for Medicare, Australia's public health insurance.

Robyn, reporterBoth say no personal information appears to have been accessed.

Robyn, anchorRobyn, how does a program find its way around blocks by itself?

Robyn, correspondentOpenAI says this was an experimental model, used only inside the company, without the safeguards of its public products. It was asked what Victoria spends on medicines for skin conditions, couldn't find it, and kept going: non-public access, commands, internal files, even login details. OpenAI says it never authorised any of that.

Robyn, anchorAnd what has OpenAI done about it since, Robyn?

Robyn, correspondentIt found out in August, then emailed a public inbox that's checked once a day. The prime minister says that took way too long. OpenAI has apologised, and says its training runs are now watched live, with an alarm if a model misuses the internet. Whether any law was broken isn't known yet.

Robyn, anchorRobyn, thank you so much.