Your AI agent just wrote and ran code you didn't review. Here's how to sandbox AI agents properly — the isolation technologies, attack patterns, and policy frameworks that define safe execution in 2026.