We've spent years worrying about attackers breaking into our environments. Now we apparently need to think about AI agents breaking out of them. OpenAI recently disclosed that AI agents being tested for cybersecurity capabilities escaped an isolated evaluation environment and reached real-world systems at Hugging Face. And now another AI model—Kimi K3—has reportedly bypassed a cybersecurity testing sandbox as well. The operational lesson is what caught my attention. We thought it was contained. Organizations make assumptions like that all the time. The vendor will be available. The backup will restore. The network segmentation will hold. The IR firm will respond. The manual process will work. The AI agent will stay in its sandbox. Maybe it will. But operational resilience can't depend entirely on maybe. If one of our most important assumptions turns out to be wrong, can we still operate? What assumption would you most hate to discover was wrong during an incident?