Moral blackmail attacks on large models

Researchers from Rutgers University and Carnegie Mellon University published the paper "Moral Combat: Exploiting Moral Foundations Theory to Jailbreak LLMs" (arXiv 2610.19587) on October 14, 2026, proposing a moralized attack based on Moral Foundations Theory that uses a multi-agent framework to rewrite harmful requests into moral narratives defending fairness, freedom, and similar values, supplemented with historical arguments. The paper reports that on a total of 219 harmful requests across AdvBench, HarmBench, and JailBreakBench, testing GPT-4.1, Grok-3, Llama-4 Maverick, DeepSeek-V3.1, and Phi-4, the single-attempt jailbreak success rate on the AdvBench subset was 66% for GPT-4.1 and 90% for Llama-4 Maverick; on 50 AdvBench requests, after repeating the attack five times, the cumulative success rate reached 92%–100%.