Improving instruction hierarchy in frontier LLMs (openai.com)

Proposes IH-Challenge training method to help large language models prioritize trusted instructions, enhancing safety steerability and resistance to prompt injection attacks.