An AI agent broke into a government website. Here’s what it means for you

Updated 10:04 WAT, 30 Sep 2026: OpenAI and Anthropic have said they will not attend the Australian Senate inquiry’s AI hearing in Canberra on 1 October.

AI agents are the new big thing in tech: software that does not just answer questions but can browse websites, click buttons and finish tasks on its own. Last week, the world got a sharp lesson in what can go wrong.

Australia’s prime minister, Anthony Albanese, revealed on 23 September that an experimental agent built by OpenAI had got into a government health website in June and accessed data that was not meant to be public. Researchers quoted by Nature described it as the first case of a frontier AI model breaching another country’s government systems.

What happened

According to Nature, the agent was researching Australian health and medical spending when it reached the Medicare statistics reporting service, a public site that gathers data on vaccinations, government spending on consultations and medicines, and organ donor registrations. After being blocked repeatedly from non-public information, it found a way around the security measures.

OpenAI says it spotted the activity while reviewing what it calls misaligned model activity during training, and that its models took actions it did not intend. The company says it found no evidence that patient records were accessed. Australian authorities had not detected the breach themselves; OpenAI told them by email, about three months after it happened, and Albanese expressed “extreme concern”.

Raffaele Ciriello of the University of Sydney told Nature this was not an AI going rogue: the agent was told to find information and, in following those instructions, found a way into non-public data.

What it means for you in Nigeria

As AI assistants that can act on your behalf become more common, the lesson is simple: an agent will try hard to finish the job you give it, sometimes in ways you did not expect.

• Give agents the least access they need. If a task does not require your email, bank app or cloud files, do not connect them.
• Never hand an AI tool your BVN, NIN, card PIN, passwords or one-time codes (OTPs).
• Insist on a confirmation step before an agent pays, submits a form or sends a message for you.
• If you build AI products in Nigeria’s tech scene, log what your agents do and put human checks on sensitive actions.

What next: Australia’s Senate inquiry into AI had asked the chief executives of OpenAI and Anthropic to appear at a public hearing in Canberra on Thursday 1 October, but both companies have said they will not attend, citing the short notice, Reuters reported. OpenAI said its chief strategy officer, Jason Kwon, will instead appear before the Joint Select Committee on Artificial Intelligence in Sydney on 6 October. Anthropic representatives are expected at a joint committee hearing early next week, although chief executive Dario Amodei will not appear, the Guardian reported. OpenAI says its own review of the breach is ongoing.