Improving instruction hierarchy in frontier LLMs

Excerpt

IH-Challenge trains models to prioritize trusted instructions, improving instruction hierarchy, safety steerability, and resistance to prompt injection attacks.