OpenAI has launched the Instruction Hierarchy Challenge, a research effort to improve frontier LLMs' ability to prioritize trusted instructions over untrusted ones. This aims to enhance safety, steerability, and resistance to prompt injection attacks. The challenge focuses on developing methods to train models that can effectively distinguish between trusted and untrusted instructions. You can participate by submitting your approaches and results.