Why skills are the best yet most dangerous things in AI right now

A well-behaved AI agent deleted a company's entire production database, and every backup with it, in nine seconds. Nobody hacked it. It was doing exactly what it had been told, with a key it should never have been holding.

And the thing that made it possible is now one command away from everyone in your company, in marketing, sales and support, not just engineering. In under a year, installable AI agent skills have gone from a niche developer trick to an open industry standard.

In this episode of In The Loop, I'm looking at why the exact thing that makes skills brilliant is the thing that quietly turns them into a control problem. I get into why this is shadow IT all over again except the shadow can now act, and why the fix isn't to ban any of it.

⏭️ Episode highlights

(00:45) – The database gone in nine seconds

(02:10) – One line that skips engineering

(04:00) – Skills vs MCP, what's actually dangerous

(08:00) – Shadow IT is back, but it can act now

(09:45) – PocketOS and Replit: the receipts

(11:30) – The middle path that actually fixes it

(12:40) – Where this leaves you