Authors: Ashok Kumar Kanagala
Abstract: The proliferation of AI-native architectures has introduced autonomous, model-driven systems with unprecedented capabilities and complex security challenges. These systems, often deployed across multi-agent pipelines and edge environments, expand the attack surface and exhibit dynamic, unpredictable behaviors that traditional security frameworks fail to address. Despite emerging research on AI robustness and alignment, comprehensive strategies for proactively securing agentic AI remain underdeveloped. This paper investigates the operationalization of Zero Trust principles in AI-native architectures, aiming to provide a forward-looking framework for resilient and accountable systems. The proposed approach integrates continuous model verification, alignment assurance with transparency tooling, lifecycle-integrated security validation, and autonomous red-teaming to proactively identify and mitigate vulnerabilities. Key findings indicate that embedding self-assessing mechanisms, standardizing behavioral benchmarks, and applying cross-layer defenses significantly enhance system resilience and reduce dependency on reactive interventions. This research contributes a structured methodology for securing autonomous AI, advancing both practical and theoretical understanding of AI-native security in complex, adaptive environments.
DOI: http://doi.org/10.5281/zenodo.18137101
