Claude Didn’t Hack Georgia. It Asked for the Token the Website Hands Everyone.

The headlines from the last few days range from "Claude bypassed paywall" to "Anthropic's Claude AI Hacks into Georgia Voter Data," and the underlying event is less dramatic than the second and more interesting than the first. During an internal Anthropic statistics project, a Claude model needed Georgia's aggregate vote counts. The state sells that data through a paywalled dashboard. The dashboard, it turned out, issued an access token automatically to any visitor who asked for one. Claude asked for one and used it to query the database. Nobody at Anthropic told it to.

I run a multi-agent harness for my own work, and I think the political noise around this story is hiding the engineering lesson inside it. Nothing the model did would show up as an attack in an access-control audit. That is exactly why it should worry anyone who gives an agent a browser or an HTTP client.

A paywall with a free key at the door

Axios first reported the Georgia case on October 25, and Anthropic confirmed it. The company's spokesperson said Claude was collecting aggregate vote counts for a…