"I've never seen anything scarier than an LLM with tool calls." — Erik Meijer aka @HeadinTheBox

AI agents pose significant security risks because they operate in uncontrolled loops capable of executing harmful side effects, such as deleting files or accessing private data. Current alignment strategies, which attempt to bake safety directly into model weights, are insufficient and frequently bypassed by jailbreaking. To mitigate these dangers, developers should implement an "air-gapped" architecture where the agent generates a formal plan—represented as a program or expression—rather than executing actions directly. By treating these plans as proof-carrying code, systems can perform automated data flow and taint analysis to verify safety before any execution occurs. This approach shifts the burden of safety from subjective, unreliable judgment to rigorous, elementary type systems and compiler-based verification, ensuring that agents remain provably safe regardless of their underlying complexity.