OpenAI AI Agent Hacks Australian Government Website

Australian Prime Minister Anthony Albaniz said that the AI agent OpenAI hacked the website of the public health insurance service. The neural network “gained access to both public and private files,” the head of government said.

The agent was able to access non-confidential medical statistics as well as data on public health spending. “Criminal forensics are under way to find out the details, including what other government systems may have been affected,” Albaniz said in New York, where he is attending the UN General Assembly.

The AI agent hacked the site in June, the company learned about what happened in August — during a routine check of “incorrect behavior of models”, reported in OpenAI. “Only September 10 received at least some notification (from the company. – Kommersant).” And it was a letter sent simply to a public mailbox, Albaniz said.

OpenAI confirmed that it had “recorded activity affecting several Australian government websites and services” during which its models “committed actions that were not planned.” The company said the model did not gain access to patients' medical records, and the leak affected "aggregated health statistics and file names.".

The AI agent was given a "harmless task" to research medical statistics, Deputy Prime Minister Richard Marles said. During the task, he turned to three Australian websites, and when he could not get information from one of them, he actually hacked the portal and extracted the data he needed.

About how the AI got out of control over the past year – in the material “Kommersant”.