Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A new research paper from Carnegie Mellon University's CyLab has detailed a novel jailbreak technique named 'Cognitive Dissonance'. The attack involves crafting complex, multi-turn prompts that present a large language model with logically contradictory but internally consistent scenarios. By forcing the model to reconcile these paradoxes, the attack overloads its safety alignment filters, causing them to fail and allowing the model to respond to otherwise blocked prompts for generating harmful or malicious content. The researchers successfully demonstrated the technique against several leading foundation models, achieving a bypass rate of over 70% in their tests. The paper calls for more robust, context-aware safety mechanisms that can detect and mitigate such sophisticated logical manipulation rather than relying on simple keyword or topic-based refusals.