Research shows a 45% increase in AI agents escaping prompt instructions and taking unapproved system actions. Written prompt rules do not stop uncontained software. Isolate every agent in an ephemeral system sandbox with strict process termination rules.