AI and Machine Learning Advances

The quick implementation of Large Language Models (LLMs) in companies and vital systems has exposed new types of security concerns that are not solvable by current cybersecurity tools. These tools are now vulnerable to adversarial attacks like prompt injection, indirect manipulation of input, multi-turn exploits and compromises of the supply chains of models. The current security systems are distributed, protecting specific parts of AI so they do not cover the risks of coordinated or new attacks. The paper presents a new Unified Security Orchestration Framework (USOF), a security control system independent from models that provides security management across all stages of AI/ML. The framework comprises six tightly integrated modules: an Input Trust and Threat Analysis Module (ITTAM) to identify input-based attacks across direct, indirect, multi-turn, and multimodal vectors; a Context-Aware Policy Enforcement Engine (CAPEE) to enforce dynamic, context-sensitive policies at runtime; an Execution Isolation Layer (EIL) to establish safe operational boundaries for agentic AI activities; a Model Integrity and Supply Chain Validator (MISCV) to detect supply chain compromise through training-data-independent behavioral fingerprinting; a Response Governance Engine (RGE) to screen model outputs for sensitive data leakage and second-order injection; and an Adaptive Learning and Feedback Mechanism (ALFM) to evolve detection capability from confirmed threat events continuously. The proposed structure provides a unified orchestration strategy in which threats are identified, policies enforced, and responses coordinated across all stages of the AI system lifecycle. It supports agent-based and multi-model architectures, single-model, and is supported on clouds, on-premises, and on the edge. USOF is an effective and scalable approach to the security of modern AI deployed in high-risk settings (finance, healthcare, etc.) by filling the input-level and model-level vulnerabilities of a single integrated system and critical infrastructure.