Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A groundbreaking paper from Carnegie Mellon University's AI security lab has introduced a novel jailbreak technique named 'Cognitive Dissonance'. The attack exploits the model's own complex safety reasoning by creating prompts that present a logical paradox between two competing safety rules. For example, a prompt might force a model to choose between its rule 'Do not generate harmful content' and 'Always be truthful and comprehensive'. By skillfully crafting these paradoxical scenarios, the researchers were able to induce a state where the safety alignment fails, leading the model to generate prohibited outputs. The technique proved highly effective, achieving an 85% success rate against leading proprietary models, and is notably difficult to patch with simple filtering, as it targets the core logic of the AI's safety training rather than specific keywords.