The Coastal Virginia Center for Cyber Innovation Cybersecurity (COVA CCI), Southeastern Virginia’s node of the Commonwealth Cyber Initiative, awarded $480,000 for six cybersecurity research projects focused on Agentic AI Cybersecurity and Security for Agentic AI Systems. Teams from Christopher Newport University, Old Dominion University and William & Mary submitted proposals in response to the COVA CCI request for proposals.
This call for proposals asked researchers to submit proposals that advance the state of the art of agentic artificial intelligence (AI) for cybersecurity and security for Agentic AI systems. Agentic AI refers to autonomous AI systems that can act independently to achieve goals. The core concept of agentic AI is the use of AI agents to perform automated tasks with limited human intervention. These systems plan, execute, and adapt to complex, multi-step tasks without continuous oversight by using a combination of large language models, reasoning, and action-taking capabilities. “As we enter the era of the Internet of Agents, security must be the foundational architecture, not an afterthought. By funding these strategic grants, we are empowering our researchers to build the next generation of autonomous systems that act with integrity, transparency, and resilience. “ said Daniel Takabi, Director of COVA CCI.
Objectives of this call include:
Project team: Dr. Phillipe Giabbanelli, Old Dominion University and Dr. Alexander Nwala, William & Mary.
https://www.odu.edu/directory/philippe-giabbanelli
https://cdsp.wm.edu/data-science/people/nwala-alexander.php
Project Abstract: Agentic AI systems use autonomous agents powered by large language models (LLMs) to plan, act, and interact with limited human oversight. While such systems are increasingly deployed in operational and customer-facing settings, their design remains largely unstructured: agents are composed through ad hoc workflows with little ability to verify whether their behaviors and interactions are coherent, safe, or secure before deployment. Existing agentic AI platforms rely on runtime guardrails or safety checks after agents have acted. In open, adversarial contexts, this post-hoc approach risks cascading or irreversible outcomes. Emerging solutions for verification operate over completed workflow, so designers must have finished their system before checking. In contrast, we propose a verifiable, agent-centric architecture that enables security analysis of agentic AI systems during early stages of the design. Building research in semantically verified agent-based modeling and empirical research on coordinated and inauthentic online behavior, we introduce behavioral building blocks that explicitly encode the semantics of agent actions and interactions. These building blocks support automated verification of semantic compatibility, enabling the detection of unsafe or incompatible agent compositions before a system is executed. We will design, implement, and evaluate this architecture through 3 case studies that cover realistic cybersecurity scenarios with established performance baselines as well as design-time failures, such as incompatible behavioral requests and unsafe chains of actions.
Project Team. Dr. Jonathan Takeshita, Old Dominion University and Dr. Qun Li, William & Mary
https://www.odu.edu/directory/jonathan-takeshita
Project Abstract. Agentic AI systems, autonomous systems that plan, reason, and act with limited human oversight, are emerging as a foundation for next-generation cybersecurity and decision-support workflows. Protecting agent plans, memory, credentials, and decision integrity is particularly challenging in adversarial or untrusted execution environments. Trusted Execution Environments (TEEs), such as Intel SGX, offer strong hardware-enforced confidentiality and integrity guarantees, but their scalability limitations conflict with the computational and memory demands of agentic AI. Recent TEEs trade security for scalability, an unacceptable compromise for high-assurance autonomous agents. This project builds on our prior work that augments SGX with cryptographically protected use of untrusted memory, preserving a strong adversary model while improving scalability. We propose to investigate how these techniques can be applied to agentic AI systems, enabling secure execution of autonomous planning, persistent agent state, and tool-mediated actions without weakening SGX’s security guarantees. Using a deliberately scoped, one-year effort, we will develop a proof-of-concept secure agentic workflow that integrates enclave-resident control logic with protected external memory, authenticated tool invocation, and remote attestation. The results will provide empirical insight into the feasibility, performance, and security tradeoffs of deploying trustworthy agentic AI in hostile environments, laying the groundwork for future large-scale research and deployment.
Project Team. Dr. Qing Tang, Old Dominion University and Dr. Ye Gao, William & Mary
https://sites.google.com/view/qingtang
https://www.wm.edu/as/computerscience/people/gao_ye.php
Project Abstract. Connected and automated transportation systems increasingly rely on agentic AI to integrate heterogeneous sensing, communication, and traffic state information for perception and decision-making. These systems are vulnerable to cyber-physical attacks and sensing degradation—such as camera blinding, LiDAR spoofing or removal, and falsified V2X messages—that can compromise perception integrity and propagate safety and mobility risks through traffic-flow dynamics. Existing anomaly and intrusion detection approaches, which largely operate at the single-modality or component level, are ill-suited to detect such coordinated cross modal and cross-state integrity failures. This project proposes an agentic, evidence driven cyber-physical security framework that operationalizes security through perception integrity, defined as the correctness and reliability of the fused world model under adversarial or degraded conditions. The core premise is that multimodal consistency and physical feasibility provide attack-agnostic integrity signals, manifested through violations in cross-modal alignment, predictive reconstruction fidelity, and physics-feasibility and traffic-dynamics constraints. The research will (i) construct a multi-modal cyber-physical integrity and threat dataset, (ii) develop a multi-modal integrity evidence generation framework, and (iii) design an autonomous agentic AI reasoning module that maintains evolving integrity beliefs, attributes likely cyber-physical threats through multi-step evidence-grounded reasoning and recommends representative mitigation responses. The proposed work directly aligns with the COVA CCI mission by advancing cybersecurity, trustworthy autonomy, and resilience in safety-critical cyber-physical systems.
Project Team. Dr. Lusi Li, Old Dominion University, Dr. Shuai Hao, Old Dominion University, and Dr. Yan Lu, Christopher Newport University
https://sites.google.com/view/lusili/
https://www.odu.edu/directory/shuai-hao
https://cnu.edu/faculty/yanlu.html
Project Abstract. Agentic artificial intelligence (AI) systems increasingly operate as collaborative collectives, where multiple agents exchange and fuse task-relevant semantic information (e.g., embeddings, symbolic attributes, compressed scene descriptors) rather than raw data streams to achieve complex, shared objectives. Clustered semantic communication (SemCom) is the enabling technology for this collective intelligence, allowing dynamically formed groups of agents to fuse partial, heterogeneous observations into a coherent shared understanding that supports coordinated perception, planning, and action. While conventional security mechanisms can protect the bit-level channel, such as encryption and authentication, they are largely blind to attacks that manipulate meaning. Adversaries can manipulate semantic representations, induce cross-agent inconsistencies, or poison shared knowledge bases (KBs), thereby degrading collective behavior without violating traditional communication guarantees. This project addresses security challenges unique to clustered SemCom in agentic AI systems. We consider a hybrid knowledge architecture in which each agent maintains a private KB while shared knowledge evolves asynchronously through clustered semantic exchanges and optional system-level aggregation. We will (i) systematically model adversarial threats targeting semantic representations, fusion processes, KB updates, and LMM-based reasoning loops, and (ii) develop a semantics-aware security framework that enforces semantic consistency, KB integrity, and decision robustness. By grounding security mechanisms in meaning rather than bits, the proposed framework will establish foundational principles and practical defenses for trustworthy agentic AI collectives, enabling resilient autonomous systems that can operate reliably under adversarial and uncertain conditions.
Project Team: Dr. Frank Liu, Old Dominion University and Dr. Haipeng Chen, William & Mary.
https://www.odu.edu/directory/frank-liu
https://cdsp.wm.edu/data-science/people/chen-haipeng.php
Project Abstract. The objective of this project is to develop methods and techniques that ensure the safe operation of high-stakes physical systems (such as complex high-energy accelerators) when contemporary LLMs (large language models) are deployed to agentic agents. By integrating knowledge from different domains with heterogeneous sensor data, operational context, and procedure knowledge, AI agents can provide real-time adaptive control decisions that usually exceed the capability of traditional fixed policy controls. However, safety remains a major open challenge because the current LLMs may produce seemingly confident outputs that violate physical or operational limits. This risk is amplified in high-stakes physical systems, where small reasoning errors can cause large and potentially irreversible negative consequences. Our research is anchored in a rigorous foundation and state-of-the-art frontier LLM models, with concrete use cases from high-stakes physical systems (e.g., state-of-the-art accelerators) to shape applicable and generalizable methods and techniques.
Project Team. Dr. Yue Xiao, William & Mary and Dr. Mahmoud Nazzal, Old Dominion University.
https://www.wm.edu/as/computerscience/people/xiao_yue.php
https://www.odu.edu/directory/mahmoud-nazzal
Project Abstract. LLM-powered mobile agents are increasingly used in real-world tasks such as trip planning, food ordering, and workflow automation. To perform these tasks, they interact with multiple applications and system resources, resulting in expanded and largely unexplored attack surfaces that require systematic analysis and mitigation. This project directly addresses this challenge. Mobile ecosystems introduce security challenges that are specific to their execution and communication model. Untrusted third-party software development kits (SDKs) can inject adversarial content through advertisements, notifications may carry untrusted external inputs, and cross-app interactions can induce context confusion across application boundaries. These factors render mobile LLM agents susceptible to confused-deputy attacks, in which agents are manipulated into misusing delegated privileges or inadvertently exposing sensitive information. Our prior work shows that carefully crafted adversarial inputs can induce unsafe agent behaviors, including privilege escalation, data leakage, and malware installation. Building on these findings, this project moves beyond isolated case studies to develop a systematic framework for analyzing and defending agentic AI systems in mobile environments. Specifically, the project is organized around three tightly coupled thrusts: 1 automated generation of realistic adversarial scenarios for mobile agents, 2 principled measurement and quantification of agent misbehavior, and 3 graph-based runtime defenses to detect and mitigate confused-deputy behaviors. Conducted jointly by William & Mary and Old Dominion University, the project combines complementary expertise in mobile systems security and agentic artificial intelligence. The resulting tools and defenses will strengthen the Commonwealth’s capacity in trustworthy agentic artificial intelligence.
