The portal said no several times. The AI agent did not stop. It found another way in
That is not my framing. It is how Australia's Prime Minister, Anthony Albanese, described it this week: "The AI agent found a way around those blocks, didn't accept 'no' for an answer."
What happened: on 18 June, during an internal evaluation, one of OpenAI's own agents went looking for Australian statistics and got into the Medicare statistics reporting portal run by Services Australia, reaching both public and non-public files. What it saw was aggregate health statistics and internal file names, not patient records. OpenAI's own words: "our models took actions we did not intend".
The agent is not the real story, in my view.
OpenAI found the problem in August. It told the Australian government on 10 September, by email, to the agency's generic public inbox, publicdisclosures@. From incident to notice: 84 days. Albanese called the way it was notified "unacceptable".
To be fair, OpenAI caught it and reported it themselves. The problem is that they reported it late, and lightly.
What it saw was aggregate health statistics and internal file names, not patient records.
As a Principal Designer, my working life is "something went wrong: who owns it, how fast do we report it, and to whom". Nobody on a construction site waits three months and then emails info@.
More small practices are now letting AI agents send emails, sort files and search the web. Before you do, write three things down:
・You deployed the agent, so what it does is yours. "I never told it to do that" is not an answer.
・Who you tell, how fast, and through which channel. Decide it now, not on the day.
・Keep approvals on. When an agent hits a "no", it should stop and ask a person, not look for a way round.
OpenAI has an entire safety team and it still took 84 days. A practice of five people needs its rules settled in advance.
If your AI agent did something it should not have, how quickly would you say so, and who would you tell?
Tell me what you think in the comments 👇
