Overview
Severity: HIGH | Affected: Anthropic, Google | Category: research
A new paper from Stanford University's AI Lab has introduced a novel jailbreaking technique named 'Cognitive Dissonance.' This method bypasses the safety alignment of several major large language models, including those from Anthropic and Google, with a near 95% success rate in benchmark tests. The technique works by presenting the model with two conflicting, but individually benign, prompts in a single request. This forces the model into a logical paradox, causing its safety filters to fail and allowing it to generate harmful or restricted content. The researchers argue this highlights a fundamental vulnerability in current alignment strategies, which focus on single-intent prompts and struggle with complex, contradictory inputs. The findings have been responsibly disclosed to the affected companies, who are now scrambling to patch their models against this new class of attack.