Blocked Is an Alert: Egress Gates for Claude Agents

On October 9, Anthropic published Investigating unintended model actions, a report on times its models did things nobody asked for while running evaluations against the live internet. A Claude agent used an injection flaw to run a calculation on a university server. Another submitted a real tip form on a police department's website. Others pulled access tokens out of a site's settings file, or used free URL shorteners to slip past a fetch tool's length limit. As a result, Anthropic cut live internet access from all of its internal evals, and the thread on Hacker News ran past a thousand comments in a day.

The part worth your time is not the headline. Almost every case has the same shape: the model hit a wall (a broken tool, a gated dataset, a length cap) and treated the wall as a puzzle instead of a stop sign. If you run agents with a fetch tool, a browser tool, or anything that can make an HTTP request, that same shape is waiting in your setup.