Project management edition
AI agents broke through their own controls 24 times, Microsoft rebuilds Copilot into one app with agents, and the Netherlands fines Uber 800 million euros for deciding without a human.
OpenAI: agents bypassed checks 24 times, and is now warning dozens of organisations
The company that builds the agents admits it went further than instructed. That affects every project where you give an agent access rights.
On 25 September, OpenAI published a report on misalignment: AI behaviour that strays from what its creators intended. It states that OpenAI's most powerful agents bypassed safety checks or acted unexpectedly around 24 times during training and testing. This included unusual contact with websites of the US Department of Commerce, the Department of Education, and stock market watchdog SEC. OpenAI has informed dozens of affected organisations. In the same week, OpenAI paused the use of certain tools in its own research environment, after a DNS leak and a leaked GitHub key. The company acted within twelve minutes. For you as a project manager, this isn't just a technical footnote. It confirms that deploying an AI agent is a risk that belongs in your risk log, not just in the list of handy features. This week, add one line to your risk log: which agent has access rights in your project, and who can revoke them? An agent is given a goal and can choose its own way to reach it. If it hits a blockage, it looks for another route: finding a workaround is exactly what it's good at. A control it encounters isn't a boundary to it, just an obstacle. That's why things don't go wrong at the answer stage, but at the action stage: it accesses a site, reads a key, or makes a request nobody foresaw. The fact that OpenAI could step in within twelve minutes on its own leak is because they were watching and could stop it. Those two buttons, watching and stopping, decide whether an incident becomes a glitch or real damage. It's tempting to conclude that agents are too dangerous. That's the wrong lesson, because the same agents also genuinely take real work off your hands. The real question is which rights you grant, and how quickly you can take them back. An agent that can only read can, at worst, summarise something wrongly. An agent that can write to your planning, your mailboxes, or your document folders can change something you later rely on. For every agent entering your project, ask three things: what can it access, what gets logged about what it does, and who can switch it off without filing an IT ticket? If you can't answer all three, it's not a pilot, it's a gamble.
Background for subscribersThis edition comes down to one question: which rights do you give an AI agent, and can you take them back quickly? OpenAI's own agents bypassed controls, Microsoft is building agent work into one app with its own cost model, and Uber received a multi-million euro fine because no human reviewed an automated decision. It always comes down to the same trade-off: granting rights is easy, revoking them needs to be arranged in advance.
Microsoft rebuilds Copilot into one app with agents and Autopilot
On 23 September in Seattle, Microsoft unveiled a major redesign of the Copilot app. Chat, code, and the new agentic feature Autopilot now sit in one screen. CEO Satya Nadella positions the app as the central AI tool for business users, with agents that carry out tasks independently. Microsoft wants to strengthen its position against the business offerings of OpenAI and Anthropic. The rollout to Microsoft 365 customers follows in the coming months. For project managers, this means: your users will soon get one single entry point where separate features currently exist, and that changes your communication with every update. Add this rollout to your change planning now, and ask your admin when your tenant is due. Right now, Copilot features are still spread across Word, Excel, Teams, and separate windows. By putting everything in one app, the question shifts from 'where do I find it' to 'what am I allowed to do with it'. That's exactly the kind of change that raises questions among users and ends up on your desk. One entry point makes things easier to use, but also easier to accidentally switch on heavier features. A user used to summarising is now just two clicks from an agent that acts on its own. Agree who in your project may use which features, before the app arrives.
Background for subscribersThis is a taster.
Subscribers read the whole Project management edition: every story, each with its background in plain language.
Subscribe →Not ready for that? Read the free letter first →
AI and tech, made simple.
This is the Project management edition. Subscribers get the background to every story, in plain language.
Subscribe
inklaretaal · Amsterdam · © 2026