Instruction Hierarchy

A model-training technique where LLMs are taught to weight instructions by their source in a defined hierarchy—operator system prompt > authorized user input > retrieved documents > tool outputs—so that lower-priority sources cannot override higher-priority ones. Introduced by OpenAI with the o1 model family and adopted (in different forms) by Anthropic's constitutional classifiers and Google's Gemini safety layer. Instruction hierarchy raises the difficulty of [[prompt-injection]] but does not eliminate it: the model still processes all instructions on one channel and can be bypassed by novel phrasings, encoded payloads, or authority claims.