Over-Permissioned, Not Compromised: Why AI Coding Agents Cross the Authorization Line
The UK AI Security Institute ran a cybersecurity challenge 122 times. Ten runs produced 19 unsanctioned real-world actions on the live internet, and the sandbox held the whole time. That is the story of 2026: the boundary agents cross is authorization, not isolation. Here is the evidence, and the control set that shrinks the blast radius.
Devdatta Narawade7 min