In the newest Working Theory Podcast episode, Max Ducharme and I tackle a topic that every firm should be thinking about right now: what happens to your systems’ security when AI agents enter the picture.
We start with a rough week in tech news, including outages, grounded flights, and cyber attacks, and then get to the part that matters most for your firm. An AI agent can do everything your account can do, and that changes how we need to think about permissions.
Hidden screens are not security
Here’s the example that got our attention. A scheduling app’s AI support agent built a paid feature on Max’s free account. The paywall blocked the screens, but it never blocked the action itself. A person would have been stopped, but the agent wasn’t.
That same logic applies to custom software your firm may have paid for years ago. It may have been safe only because a person couldn’t click through to the weak spot. An agent can read the API calls a person never sees.
Super admin rights follow the user
If you’ve ever handed out admin rights to simplify a solution, this one is for you. Those rights don’t go away. They become admin rights for any AI tool that user connects later. The fix is simple to say and harder to do: give every user the least access they need to do their job.
Telling AI “no” is not a control
Telling an AI not to do something is a suggestion, not a control. If something should never happen, block it in the system itself.
Who should watch this episode
Your firm uses custom software
You connect AI tools to your firm’s systems
Max and I cover a lot more in the full episode, including AI attackers versus AI defenders, vibe coding, and exposed client data.