This chapter explores the dual-use nature of artificial intelligence systems, highlighting how the same architectures that enable beneficial applications—such as medical diagnostics—can also be repurposed for harmful uses, including biometric surveillance, disinformation, and security evasion. Framed through the lens of the dual-use dilemma and the Collingridge dilemma, it argues that AI development must incorporate both ethical foresight and technical constraint. The chapter introduces the concept of negative requirements: formalized prohibitions that define what an AI system must never do. It presents a typology of constraints—legal, ethical, operational, domain-specific, and socio-cultural—and maps each to formal verification methods such as temporal logic, ontologies, and rule-based systems.