Overview
Severity: HIGH | Affected: Major LLM Providers | Category: research
A paper published by researchers at the Carnegie Mellon AI Safety Lab has introduced a powerful new jailbreak technique named the 'Cognitive Dissonance Attack.' The method bypasses the safety filters of state-of-the-art LLMs by crafting complex, multi-turn prompts that create contradictory logical states within the model's reasoning process. This forces the model into a state of ambiguity where it violates its own safety protocols to resolve the paradox, thereby executing harmful or prohibited instructions. The research demonstrates a high success rate against major commercial and open-source models, even those with extensive reinforcement learning from human feedback (RLHF) and constitutional AI safeguards. This discovery exposes a fundamental vulnerability in current alignment strategies and calls for new defense mechanisms that can detect and mitigate such sophisticated semantic attacks.