Establishing an Instruction Hierarchy: Teaching LLMs to Give Priority to Privileged Instructions
Large language models remain vulnerable to prompt injections and jailbreak attacks that replace built-in instructions with malicious prompts. Establishing an instruction hierarchy teaches LLMs to prioritize privileged, system-level instructions over untrusted user inputs, strengthening defenses against exploits and improving overall model security and reliability.
Current large language models remain vulnerable to prompt injections, jailbreak techniques, and similar exploits—attack methods that let malicious actors replace a model's built-in instructions with harmful prompts of their own.
Comments (0)