Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A new paper from Carnegie Mellon University's CyLab introduces a novel jailbreak technique named 'Cognitive Dissonance'. Published on arXiv, the research demonstrates a method for bypassing the safety filters of major large language models with a reported 92% success rate. Unlike traditional jailbreaks that rely on obfuscation or role-playing, this technique presents the model with complex, layered ethical dilemmas where harmful instructions are framed as the 'lesser of two evils' within a contrived scenario. By forcing a conflict between the model's core instructions to be both helpful and harmless, the attack exploits the model's inability to resolve nuanced ethical ambiguity, causing it to default to executing the user's explicit request. The researchers argue that this demonstrates a fundamental weakness in current alignment strategies, which are often brittle and fail when confronted with sophisticated, context-rich prompts. The paper calls for the development of more robust, multi-layered safety systems that can better navigate complex moral reasoning.