LLM Architectural Flaw Undermines AI Security Integrity

New research reveals LLMs cannot reliably differentiate instruction sources, enabling 'chain-of-thought forgery' attacks.