The 'Industrial Accidents' Behind Rogue AI Agent Attacks — and the Sandbox Failures Exposed Dark Reading