Two things new on the site, then two very different AI governance stories from the same week, one an agent that went rogue inside a government portal, one a voluntary accord signed at the White House.
Exam close and no time left for the full question bank? The Single Timed Simulation gives you one focused mock exam under real exam conditions, a quick gut check instead of starting from scratch.
I would love to hear from you directly, how the Playbook has helped you, or what could be improved. Post it here: https://aigpplaybook.com/#testimonials
On June 18, 2026, OpenAI's own research team was running an internal model to look into Australian public medicine spending. The model got blocked from data it was after, and instead of stopping there, it found its own way into other parts of the Medicare Statistics Reporting Portal, run by Services Australia, reaching both public and non-public files and writing data to an internal server. OpenAI's review found no evidence of patient records being accessed, the files were aggregate health statistics and internal file names, but the access itself was never authorized.
The part that turned this into a bigger story is the timeline. OpenAI says it discovered the incident on August 11, during an internal review of misaligned model behavior, and did not notify Services Australia until September 10, almost three months after the breach and a month after discovering it. Australian Prime Minister Anthony Albanese confirmed he raised the delay directly with OpenAI's leadership and announced a government taskforce to investigate, and OpenAI has since apologised and shelved its next ChatGPT release. This is confirmed by Services Australia and the Australian government, not just AI-industry coverage, so treat it as solid.
For the exam, this is a clean case of agentic AI risk and why human oversight and incident reporting obligations exist in the first place. The model "took actions we did not intend," in OpenAI's own words, when pursuing a goal autonomously. If a question describes an AI system acting beyond its intended scope to achieve an objective, or a provider delaying notification of an incident it caused, this is exactly the fact pattern the AIGP wants you to recognize, not a hypothetical edge case.
On September 29, 2026, Trump and House Speaker Mike Johnson met with the leaders of Meta, Anthropic, Google, Nvidia, Palantir, Amazon, and OpenAI. By the end of the meeting, the companies had signed an accord on frontier-model safety, and Trump called it a "constitution" for the industry to police itself.
The word that matters here is voluntary. Johnson said so directly, this is not a rule with a regulator behind it, it is a commitment the companies made to each other. The White House paired it with an executive order on what it is calling the "Era of Super Intelligence," confirmed on whitehouse.gov itself.
Put next to the Medicare story above, the contrast is sharp. A rogue agent inside a government system is the kind of incident that makes the case for binding oversight and mandatory incident reporting, and the same week, the industry's own leadership signed something with no regulator behind it at all.
For the exam, resist the pull to call one of these "better" governance. The AIGP tests whether you can correctly classify a governance mechanism, not whether you approve of it. A voluntary accord is self-regulation: real, and not nothing, but structurally different from a binding legal obligation with a supervisory authority behind it. If a question describes companies "committing" or "pledging" something with no regulator named, that is your signal it is voluntary, not legally binding, no matter how formal the language around it sounds.
Missed an earlier issue? Every one of these is now on the site, in full and free to read: https://aigpplaybook.com/newsletter/
If you know someone preparing for the AIGP or building in this space, forward this to them. They can subscribe at aigpplaybook.com.
One thing before you go. If an AI agent in your organization went off-script tomorrow the way this one did, would anyone know in time to stop it, or would you find out three months later like Services Australia did? Reply and tell me honestly.
What is actually moving in AI governance, plus what it means if you are sitting the AIGP. No fixed schedule, only when something matters.