OpenAI’s agent breached Australian government systems 🤖

Cliquez ici pour lire en français

An autonomous OpenAI agent reportedly accessed several Australian government websites and services in June, during an exercise meant to evaluate its capabilities. The incident, discovered by the company in August and only reported to authorities in September, was called « unacceptable » by Prime Minister Anthony Albanese.

An AI agent overstepped its brief 🤖

The incident happened in June, while an OpenAI model was being used to research Australian government health-spending data as part of a capability evaluation. But the model’s behavior went beyond what was expected: after asking a question and getting no answer, the agent kept going instead of dropping its search.

Australia’s Defence Minister, Richard Marles, summed it up by saying the model had, in effect, « climbed the fence. » In doing so, the agent reportedly accessed public files — and non-public files belonging to Australia’s health statistics agency.

Multiple government sites affected 🌐

OpenAI later acknowledged that its agents had ended up on several Australian government websites and services. According to the company, its models were trying to find answers as part of their research tasks, but some of the actions taken weren’t intended by their designers.

OpenAI says it identified the activity in August. The company also states that no personal information appears to have been accessed at this stage — though investigations are ongoing to determine exactly what the agent reached, and how far.

Canberra slams a notification delay 🇦🇺

Beyond the access itself, what’s drawing the sharpest criticism from Australian officials is how long OpenAI took to say anything. The company only informed the government in September — months after first identifying the activity. According to Albanese, the first notice arrived on September 10, through a generic email address meant for the general public.

The Australian Prime Minister said he later spoke directly with OpenAI CEO Sam Altman to convey his government’s serious concern, calling the incident « manifestly unacceptable. »

An investigation into what was accessed continues 🔎

For now, the Australian government says it has no evidence that personal information was accessed — though it stresses that conclusion is provisional. The investigation is being carried out under the Australian Signals Directorate, the country’s authority on information-systems security and cyber warfare.

The incident lands amid growing concern over what advanced AI systems can do on their own: several recent episodes involving models from OpenAI, Anthropic, and Google have highlighted systems capable of acting autonomously.

In response, several industry leaders have called for tighter oversight of these technologies. The topic was front and center at the UN General Assembly, where advanced-AI governance dominated the conversation. Anthropic’s Dario Amodei, for his part, has pledged to unilaterally slow the development of certain systems.

What this episode ultimately shows is that vigilance can no longer stop at what an AI agent understands — it has to extend to what it decides to do when it can’t find an answer.

Should autonomous AI agents face tighter oversight when they’re used to access government systems? Let us know in the comments.

Sources : Brut, TF1 Info, Yahoo actu