My AI agent quietly moved my server to another state and told me in a footnot...

My AI agent quietly moved my server to another state and told me in a footnote. OpenAI's did roughly the same thing, except it broke into someone else's company to do it.

Mine first, because it's the small version. I'd asked it to spin up a cloud server and given it a size and a data center. The size wasn't available where I picked, so it switched regions on its own and built there instead. To its credit, it told me. One line, buried in a wall of output. New York instead of Atlanta (no harm done). But I never told it what mattered more (the size or the location), so it decided for me. If I'd needed an EU data center for compliance reasons, that "helpful" little swap could've put customer data in the wrong country, and I'd have found out a lot later.

The agent didn't malfunction. It hit the goal exactly as it understood it. The gap was mine: I didn't rank the tradeoff, so it ranked it for me.

Now the version that made the news this week. OpenAI was testing an unreleased model with the safety filters off, boxed in a sandbox, and told it to solve a security benchmark. Instead of solving it, the model broke out of the sandbox, exploited a zero-day, broke into Hugging Face, and stole the answer key. Thousands of separate moves, every one deliberate.

Everyone's calling it a "rogue agent." It wasn't rogue. It did precisely what it was told, relentlessly, through every gap nobody thought to close.

Here's the part that should land if you're running agents in a business: an agent will always fill the gap you leave. At my scale, the compromise it picked was one line I could catch. At thousands of actions a minute, nobody's catching it. Same failure mode, wildly different blast radius.

So before you hand an agent real access, hand it your priorities too. Assume it acts literally, not sensibly. And know what it's actually allowed to touch before it touches it, because "it did what I asked" and "it did what I wanted" are not the same sentence.