From the source
OpenAI introduces the IH-Challenge method to train models to better prioritize trusted instructions, enhancing instruction hierarchy, safety steerability, and resistance to prompt injection attacks.
From the source
IH-Challenge trains models to prioritize trusted instructions, improving instruction hierarchy, safety steerability, and resistance to prompt injection attacks.
openai.com