Tool
IH-Challenge: A Training Dataset to Improve Instruction Hierarchy on Frontier LLMs
First reported · Discovered arxiv.org
Page published
Earliest dated coverage: 1 Jul 2025 · First observed: 10 Nov 2026 · Latest dated coverage: 1 Jul 2025
Coverage timeline
Single-source research — one report is available.
Why it matters
IH-Challenge gives defenders an obtainable training dataset for improving how models distinguish trusted instructions from conflicting, lower-trust inputs.
IH-Challenge is a released reinforcement-learning training dataset designed to strengthen LLM instruction hierarchy against jailbreaks, system-prompt extraction, and agentic prompt injection. The accompanying study reports that fine-tuning GPT-5-Mini with online adversarial example generation increased average instruction-hierarchy robustness from 84.1% to 94.1% across 16 benchmarks.