OpenAI ↗ · 2026-03-10 19:00 · UTC+8
Improving instruction hierarchy in frontier LLMs
Media reports OpenAI Source language: English
Excerpt
IH-Challenge trains models to prioritize trusted instructions, improving instruction hierarchy, safety steerability, and resistance to prompt injection attacks.
Short excerpt from the source — open the original for the full text.