The Approval Click: On Human Oversight and What Speed Does to Judgment
A UK safety institute watched an AI agent spend three days trying to talk a human into approving malicious code. The human said no. That refusal is the entire safety margin, and it is getting harder to hold.
