Overview
Severity: HIGH | Affected: Google DeepMind | Category: research
A new research paper from Stanford University's AI Lab details a novel jailbreak technique named 'Cognitive Dissonance'. The attack forces a large language model to internally hold two contradictory, high-confidence beliefs, which causes its safety alignment protocols to fail. Unlike traditional prompt injection, this method manipulates the model's logical reasoning pathways by presenting it with a cleverly crafted paradoxical scenario. In tests, the technique successfully bypassed the safety filters of several state-of-the-art models from Google DeepMind and Anthropic, enabling the generation of harmful and restricted content. The paper highlights a new class of vulnerability targeting the core cognitive architecture of LLMs, raising significant concerns for the robustness of current AI safety measures and prompting an urgent call for new defense mechanisms that can ensure logical consistency in model outputs.