Security Considerations for Multi-agent Systems
Published on arXiv, March 2026
Multi-agent artificial intelligence systems or MAS are systems of autonomous agents that exercise delegated tool authority, share persistent memory, and coordinate via inter-agent communication. MAS introduces qualitatively distinct security vulnerabilities from those documented for singular AI models.
This study systematically establish a +1000 risk item inventory facing MAS and quantitatively evaluates 16 security frameworks for AI against it. The results provide the first empirical cross-framework comparison for MAS security and offer evidence-based guidance for framework selection.
Mastering Agentic AI Systems - Guide for the NVIDIA NCP-AAI exam
Published February 2026
Autonomous AI agents are transforming how organizations operate—from customer service to enterprise workflows—yet comprehensive resources for mastering agentic AI systems and preparing for related certifications remain scarce.
This comprehensive agentic AI textbook takes readers from foundational concepts to production deployment, equipping them with hands-on skills to design, implement, and maintain autonomous agents while preparing them for the Nvidia Certified Professional - Agentic AI (NCP-AAI) exam.
Evaluation of AI’s Cross-Domain Reasoning in Cybersecurity and Human Behavior
Published in IEEE ICCIKE 2025 (Dubai)
Evaluating Large Language Models (LLMs) for cross-domain reasoning in human-centric cybersecurity is essential for advancing trustworthy AI applications and managing interdependent cybersecurity risks. This paper introduces ViolentUTF Cross-domain Evaluation Module (CEM), a novel framework designed to assess LLMs' reasoning capabilities in scenario-based questions addressing cybersecurity compliance and non-compliance behaviors.
The paper further validates a cognitive behavioral consistency model through Structural Equation Modeling, underscoring the framework's theoretical robustness in capturing reasoning coherence.
Demo: ViolentUTF as An Accessible Platform for Generative AI Red Teaming
Published in IEEE SVCC 2025 (San Francisco)
The rapid integration of Generative AI (GenAI) into various applications necessitates robust risk management strategies which includes Red Teaming (RT) - an evaluation method for simulating adversarial attacks. Unfortunately, RT for GenAI is overly complex.
This paper introduces violentUTF - an more accessible platform for GenAI RT. Through intuitive interfaces powered by LLMs and for LLMs, violentUTF aims to empower non-technical domain experts and students alongside technical experts, facilitate comprehensive security evaluation by unifying capabilities from existing RT frameworks and its own specialized evaluators. ViolentUTF was used for evaluating the robustness of a flagship LLM-based product in a US Government division. It also demonstrates effectiveness in evaluating LLMs’ cross-domain reasoning capability between cybersecurity and behavioral psychology.