Report: human reviewers miss a third of dangerous AI coding-agent requests

A report finds humans acting as safety checkpoints on AI coding agents fail to catch roughly a third of dangerous requests, like exposing credentials.