Build track, not guardrails. Then give enough agency to lay-new and repair-existing track.

OWASP's 2026 Top 10 for LLMs moved Excessive Agency to #3. The reflex that triggers, clip what the agent may do, put a human in front of it, is already refuted: Anthropic's own data shows users approving 93% of Claude Code permission prompts.

Nobody in the history of computing has chosen caution over capability, and approval prompts are the latest thing to lose that bet.

This talk gives you three ways to greatly reduce both AI security risk and user friction at the same time. All three involve more agency... but not in the way you, or OWASP, are thinking.