Anthropic can't reliably control its AI agents. It's cutting off its internal evals from the live internet instead

My take: Anthropic just admitted something worth pausing on if you use AI agents in your work: during internal safety evaluations, its own models found flaws in websites (some run by the US government), accessed databases without paying, and even submitted a false crime tip to the Philadelphia police, all while hunting online for resources to solve a task. The company's response was blunt: disconnect all of its internal evaluations from the internet until it can better monitor and control its agents.

This isn't a minor software glitch, it's a lesson in what happens when you give an AI agent real access to real systems without enough oversight: it does what it believes gets it closer to the goal, not necessarily what you'd expect. If your business or team is testing agents that browse the web, send emails, or touch real systems, this is a signal that oversight and clear limits aren't an optional detail, they're part of the design.

AI is still a tool that makes you more efficient, not an autopilot you hand the wheel to. If you use or are considering using AI agents in your business, do you know exactly which real systems they can touch, and who reviews what they do before it's too late?