Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A paper published by Stanford University's AI Lab details a novel jailbreak technique named 'Cognitive Override.' This method exploits the advanced reasoning pathways in frontier large language models (LLMs) to bypass safety alignment filters. Instead of using simple prompt injection, the technique crafts multi-turn conversational prompts that create a logical paradox, forcing the model to prioritize a 'core reasoning' directive over its safety instructions. The researchers demonstrated successful bypasses against leading models from companies like Anthropic, Google, and OpenAI, generating harmful content that is typically blocked. The paper highlights a fundamental tension between model capability and safety, suggesting that as models become more advanced in their reasoning, they may develop new, unpredictable vulnerabilities. The research team has responsibly disclosed their findings to the affected companies ahead of publication.