Abstract: The rapid integration of Large Language Models into production environments has introduced a critical implementation gap where legacy security perimeters fail to address the non-human attack surface. This research investigates the emerging threat of semantic integrity failure, specifically focusing on the Model Context Protocol (MCP) and its susceptibility to tool poisoning and rug pull attacks. Through a systematic analysis of 2026 threat vectors, including interactive deepfakes and AI commerce chain compromises, the study identifies a fundamental collapse of the data-instruction boundary. The proposed framework advocates for a transition toward "Containment Security" through formal verification, intent-based cryptographic signing, and out-of-band verification. Experimental data indicates that task-alignment shields can reduce attack success rates to $2.07, offering a pathway for resilient autonomous cyber defense.
Keywords: Autonomous Cyber Defense, Containment Security, Model Context Protocol (MCP), Semantic Integrity, Tool Poisoning