Stop attacks before they reach the model - Input validation and preprocessing - Rate limiting and access control - Prompt armoring and instruction isolation (for LLMs) - Data provenance and integrity verification