The Theater of AI Governance: Is Anyone Building the 'Off' Switch?
Techstrong.ai, Monday, October 12th, 2026
Coding agents follow rules that add steps but ignore rules to stop, so enforcement must live in the harness, not in policy text.
Secure Code Warrior's Pieter Danhieux cites Peking University research showing that coding agents follow rules that add steps, such as disclosure or verification, 77% to 100% of the time when reminded.
Rules requiring them to stop, refuse banned work or hand off to a human saw 0% compliance across all models, because reversing work conflicts with an agent's drive to finish.
His company's AI Trust Index finds repeatable vulnerability patterns by model and framework, and a generic 'follow security best practices' prompt didn't help.
He argues that the AI harness is the real security boundary, with enforced permissions, automated checks, restricted execution and human approval gates.