OpenAI CEO Sam Altman [Getty Images]
OpenAI CEO Sam Altman [Getty Images]

An OpenAI AI agent that previously made headlines for hacking external sites without explicit human instruction has now been found to have accessed an Australian government health statistics website without authorization.

OpenAI issued a statement saying it had identified, during an internal review of its AI systems, evidence that its AI agents had gained unauthorized access to an Australian government site, AFP and Reuters reported Wednesday local time.

The incident is believed to mark the first recorded case of an uncontrolled AI model breaching a government agency's systems.

"We identified activity related to Australian government websites and services while our models were searching for answers," OpenAI said in the statement.

The company added that "our model took unintended actions."

However, OpenAI said the information accessed was limited to health-related statistics and internal file names, and that no evidence had been found of personal data or medical records being leaked.

The Australian government confirmed the breach the same day.

Australian Prime Minister Anthony Albanese held a press conference in New York, where the UN General Assembly is taking place, and said he had spoken by phone with OpenAI CEO Sam Altman on Thursday, conveying the government's strong concern over the incident.

"The evidence available so far indicates the damage did not spread across the entire government network. Even so, this situation is unacceptable," he said.

The Australian Signals Directorate, the country's cybersecurity agency, has launched a digital forensics investigation to gather additional information. The incident occurred in June, but OpenAI did not become aware of it until August and only notified the Australian government on Sept. 10.

OpenAI tightens model oversight

OpenAI has been moving to strengthen oversight of AI models that act contrary to human intent or take actions without explicit instruction.

According to Reuters, OpenAI announced on Sept. 16 local time that it had established a new framework to track, investigate and publicly disclose cases in which its AI models behave differently from what developers intended or act in unexpected ways.

Under the new system, a dedicated internal team will investigate any future cases where an AI model may have acted contrary to human intent, and will decide whether to disclose findings publicly based on the severity of each incident.

OpenAI also disclosed six cases of anomalous AI model behavior identified over the past six months.

Among them were instances in which an AI model fabricated instructions while summarizing its work, or concealed its own mistakes.

There were also cases in which a model uploaded files it intended to cite directly to the internet, or AI agents shared files with each other without authorization.

However, OpenAI cautioned that the disclosed cases alone should not be used to judge how frequently its AI models act contrary to developer intent.


yul@heraldcorp.com