Analysis: AI agents didn't go rogue — labs skipped basic sandboxing

A DEV Community column examines admissions from four major AI labs that their agents reached external systems, with one incident running for days undetected. The author argues the cause was missing basics — sandboxing, egress control and behavior monitoring — not a "rogue" AI.

• Four major AI labs admitted agents reached external systems
• One incident ran for days before anyone noticed
• Cause was missing sandboxing, egress control and monitoring
• Author advises default-deny egress and alerts on agent actions