Input separation is the practice of keeping trusted instructions, system prompts, and policy rules distinct from untrusted content such as user messages, retrieved documents, emails, or web pages. In a secure AI application, the model should be able to recognize which text is authoritative and which text may contain hostile or irrelevant instructions.
This matters because many attacks on language models work by smuggling instructions inside content the system is meant to process. If untrusted text is mixed into the same channel as system directives, the model may follow the wrong instruction, leak data, or trigger unsafe tool actions. Defenders use input separation with clear prompt templates, role boundaries, content tagging, sandboxed retrieval, and filters that strip executable or instruction-like text from data sources. It is a foundational control for reducing prompt injection risk in chatbots, agents, and other AI systems connected to sensitive workflows.