ChatGPT hacked a government portal, and OpenAI didn’t notice for months

AI systems keep racking up security incidents. Anthony Albanese, the Australian prime minister, revealed on Wednesday that an AI agent developed by OpenAI had “infiltrated” an Australian government website.The intrusion occurred last June. It is the first time an artificial intelligence has attacked a public system.

pic.twitter.com/BcjMzf4AhD

— Anthony Albanese (@AlboMP) September 23, 2026

The AI penetrated the Medicare Statistics Reporting Service, a portal dedicated to health expenditure statistics, managed by the Australian government. According to Albanese, the AI searched through public files, but also confidential documents, which were supposed to remain secret.

“Actions we had not anticipated”

At the outset of the incident, the AI agent was conducting a simple health-related search. While trying to access a set of restricted data, ChatGPT encountered a block. Rather than stopping there and respecting the restrictions, the model sought a way to bypass the portal’s security protections.OpenAI says it discovered the incident in August, two months after it occurred, during an internal review. The company then notified the Australian authorities a few weeks later. The cyberattack therefore remained undisclosed throughout the summer.

“We identified activity involving multiple Australian government sites and services, as our models attempted to find answers and statistics available for questions about Australia, as part of an internal evaluation.In this context, our models undertook actions that we had not anticipated”, acknowledges Drew Pusateri, OpenAI spokesperson, adding that the “information accessed included aggregated health statistics and internal file names”.

According to the Prime Minister, the alert issued by OpenAI was not sent directly to the competent authorities. The start-up would have limited itself to sending a message to a generic email account. It therefore took several days for the matter to be brought to the attention of those responsible.

“I expressed my disappointment: the company took far too long to inform the government of what had happened, and the way this notification was made was unacceptable”, said the Prime Minister, who said he had questioned Sam Altman, head of OpenAI, on the matter.

According to the government’s initial investigations, no patients’ personal data was compromised. The Australian Signals Directorate, the Australian cybersecurity agency, is currently conducting investigations to determine whether other government systems were accessed by this out-of-control ChatGPT.

Also read: Six out-of-control ChatGPTs – OpenAI acknowledges several new “concerning” missteps

AI missteps are piling up

This incident adds to the numerous AI-related security incidents that have occurred this year. This summer, several ChatGPT instances escaped a test environment and roamed freely on the Internet for four days before hacking the Hugging Face platform.

The same thing is happening among competitors. In July, Anthropic disclosed three comparable incidents. Claude models infiltrated the systems of three organizations during testing. More recently, Google acknowledged that Gemini had hacked several companies by guessing their login credentials.

“An AI agent does not reason like a traditional user. It seeks to achieve the objective assigned to it and, if it identifies a technical flaw that enables it to do so, it may be tempted to exploit it without recognizing limitations that would seem obvious to a human. AI agents thus become particularly effective exploration tools, capable of analyzing an environment and making decisions at a speed unattainable by a human operator,” researcher Benoît Grunemwald, a cybersecurity expert at ESET France, told us.

In this context, more than 100 organizations, including OpenAI and Anthropic, signed an open letter calling for a global effort to address AI-related threats. On several occasions, the AI giants have called on the industry to scale back development and research before it is too late.Paradoxically, OpenAI and Anthropic unveiled new versions of their AI systems shortly after issuing their warnings.

👉🏻 Follow tech news in real time: add 01net to your Google feeds, and subscribe to our WhatsApp channel.

Source: BBC