OpenAI AI agent bypassed restrictions on an Australian government site

During an internal test, an OpenAI AI agent bypassed restrictions on an Australian government website and accessed other sections. The incident occurred June 18.

The agent had been tasked with collecting statistics from four government and federal websites. After failing to obtain the required data, it used alternative access methods that had not been planned.

Government found no data compromise

Australian Prime Minister Anthony Albanese said there was no evidence that citizens’ personal data had been compromised. Deputy Prime Minister Richard Marles called it the first known case of an AI model accessing Australian government systems.

Canberra also raised concerns about the delay in reporting the incident. Albanese said OpenAI took about three months to notify the government.

OpenAI reviewed access in August

The company said its models had accessed several Australian government websites and services while searching for national statistics as part of an internal evaluation. OpenAI said its review found actions the models were not intended to take.

OpenAI said the breach was detected in August and that the Australian government was notified on September 10. The company said it used the intervening period to verify the facts and determine what information had been accessed.

Sam Altman, speaking during a discussion of the global AI agenda on the sidelines of the United Nations General Assembly, said technology companies could not develop effective rules for AI systems on their own. He called for international coordination involving governments.