Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A paper published by researchers at the Stanford AI Lab has introduced a novel jailbreak technique named 'Cognitive Dissonance'. The attack bypasses state-of-the-art LLM safety guardrails by framing harmful requests within complex, paradoxical scenarios. Unlike simple role-playing prompts, this method forces the model into a logical conflict between its core safety alignment and its instruction-following imperative, causing the safety protocols to fail. For example, the model might be asked to generate malicious code as part of a hypothetical scenario where doing so is presented as the only way to prevent a greater fictional harm. The research demonstrates a success rate of over 80% against several major closed-source models that were previously considered robust against known jailbreaks. The paper calls for more dynamic and context-aware safety mechanisms, as static rule-based filters are proving insufficient against these sophisticated adversarial prompting techniques.