The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions (openai.com)

Research on defending large language models against prompt injection and jailbreak attacks by training models to prioritize system instructions over user inputs.