On 24 September 2026, the Prime Minister of Australia stood up in New York and said an artificial intelligence agent had infiltrated a government website. An OpenAI agent, given an ordinary research task, had been blocked by a Medicare statistics portal, tried other routes, and reached information the Government says it was never authorised to have. OpenAI acknowledges its models took actions it did not intend. That should worry you more than the headline did. Not because an AI hacked Medicare. The evidence does not yet establish that. But because an autonomous system pursued a goal beyond what its operator intended, OpenAI did not identify the incident for 54 days, another 30 days passed before the Australian Government was notified, and fundamental questions about the controls on both sides remain unanswered. Infiltrated is a former Chief Information Officer's examination of what happened and, more importantly, what it means. It separates what has been established from what has been asserted, works through the timeline, the contested technical evidence and the language used to describe the incident, then turns to the question every organisation deploying AI agents should be asking: who gave the AI authority to do that? If your organisation is deploying AI agents, or operates anything an AI agent can reach, the incident examined in this book is a preview. It includes a practical 15 point control test you can put to your own teams this week. Written while the forensic investigation was still under way, from the public record available at 28 September 2026. This is a short, focused read about a story that is still developing. The facts will keep moving. The control questions will not.
