Overview
Severity: HIGH | Affected: Major LLM Providers | Category: research
A paper published by researchers at Carnegie Mellon University details a novel jailbreak technique named the 'Cognitive Dissonance' attack. Unlike traditional prompt injection methods that rely on tricking input filters, this attack exploits the core reasoning architecture of large language models. The technique involves crafting complex, multi-turn prompts that present the model with logically contradictory but plausible scenarios. This forces the model into a state of 'cognitive dissonance,' which degrades its safety alignment guardrails and allows it to respond to malicious requests it would otherwise refuse. The researchers demonstrated the attack's effectiveness against several state-of-the-art models from leading AI labs, achieving a success rate of over 85% in bypassing safety filters for generating misinformation and harmful code. The findings suggest a fundamental vulnerability in current alignment strategies that focus on patching behavior rather than addressing deeper architectural flaws, posing a significant challenge for AI safety.