OpenAI teaches models to ignore untrusted instructions
OpenAI just released IH-Challenge, an open reinforcement learning dataset designed to enforce a strict instruction hierarchy (system > developer > user > tool) in frontier LLMs. Training internal model GPT-5 Mini-R on these tasks significantly boosted resistance to prompt injection attacks and safety steerability without triggering annoying over-refusals. Why it matters: As autonomous agents ingest messy web data and execute tool outputs, prioritizing trusted instructions is the only way to keep them from going rogue.