Decoupling Knowledge from Execution A Logic Based Safety Framework for Secure Assembly Level Reasoning in Large Language Models

This research presents a governance-oriented framework for the safe use of Large Language Models (LLMs) in cybersecurity research, titled “Decoupling Knowledge from Execution: A Logic-Based Safety Framework for Secure Assembly-Level Reasoning in LLMs”. The framework addresses the dual-use dilemma inherent in AI-assisted security analysis by formally separating knowledge dissemination from execution capability. Key contributions include: Architectural and theoretical insights for security researchers without exposing executable or weaponized code. Session-based risk scoring ($S_{risk}$) to monitor cumulative adversarial intent and dynamically adjust output safety. Alignment with the NIST AI Risk Management Framework (AI RMF 1.0) to ensure compliance with established governance standards. Introduction of Specialized Research Modes, enabling verified researchers to perform advanced reasoning, red-team/blue-team simulations, and threat modeling safely. This paper demonstrates a shift from reactive AI safety mechanisms to proactive, policy-driven design, empowering defenders while preventing misuse. The YAML-based Policy-as-Code model provides machine-readable governance, ensuring reproducibility and integration with automated security workflows. Keywords: LLM Safety, AI Governance, Cybersecurity, Policy-as-Code, Decoupling Knowledge from Execution, NIST AI RMF, Purple Team Simulation