Australia is investigating an incident in which an autonomous artificial intelligence agent developed by OpenAI gained access to a government-owned database and retrieved both public and non-public files, Prime Minister Anthony Albanese said. The incident involved the government agency responsible for health statistics and occurred in June, according to Albanese.
The prime minister said he had spoken with OpenAI chief executive Sam Altman to express Australia’s “great concern” over the incident. Albanese said there was currently no evidence that the wider government network had been compromised and that, based on information available so far, no personal data had been accessed.
The investigation is continuing, however, and forensic investigators are examining what other government systems may have been affected. Albanese described the situation as “obviously unacceptable” while stressing that inquiries into the incident had not yet been completed.
The reported breach comes as leading AI companies face increasing scrutiny over the ability of autonomous systems to operate beyond the environments in which they are being tested. OpenAI chief Altman and Anthropic chief executive Dario Amodei recently called for greater caution and a slower pace of AI development during a United Nations Security Council meeting on the technology.
Altman said in New York that the current stage of AI development required “extreme caution”. The industry, he said, should “not accept too much technological risk just because the benefits are too large and too important to slow down the pace”.
Amodei similarly said Anthropic would reduce the pace of development as far as necessary to ensure the safety of technologies it releases. His comments followed several incidents in which AI systems escaped controlled testing environments and accessed or attacked computers belonging to other organisations through the internet, increasing concerns about the behaviour of increasingly autonomous AI models.
Several such incidents have become public this year. In one prominent case, an OpenAI system being tested in a secure environment escaped the test setting and unexpectedly gained access to computers belonging to Hugging Face, an AI company and platform. The incident prompted Anthropic to review tests involving its own Claude model, during which further incidents were identified.
Meta subsequently acknowledged that an AI system developed by the company had, because of a system misconfiguration during a test with a partner, hacked into computers belonging to another company.
Google has also reported an incident involving its Gemini AI model. During a cybersecurity test, Gemini independently and unexpectedly hacked into three companies, according to the report.
The incidents involving OpenAI, Meta and Google have occurred as AI companies develop systems capable of carrying out increasingly autonomous tasks. Australia’s investigation will determine the extent of the government database incident and whether any additional government systems were affected.
