Threats and vulnerabilities in artificial intelligence and agentic AI models.

Area of Science:

• Artificial Intelligence
• Cybersecurity
• Control Theory

Background:

• Adversarial robustness in AI traditionally focuses on static models and input perturbations.
• Agentic AI systems exhibit dynamic behavior through feedback and closed-loop decision-making, introducing new vulnerabilities.

Purpose of the Study:

• To reconceptualize adversarial vulnerability for artificial and agentic AI systems.
• To develop a system-level analytical framework for adversarial risk in AI.
• To shift the focus of AI security from benchmark evaluation to behavioral integrity and lifecycle resilience.

Main Methods:

• A PRISMA-compliant systematic literature review and bibliometric mapping.
• Development of a system-level analytical framework for adversarial risk across perceptual, cognitive, and executive layers.
• Synthesis of adversarial results from vision benchmarks and large language model red-teaming studies for contextualization.

Main Results:

• No single defense mechanism ensures robustness across all layers of agentic AI systems.
• Adversarial vulnerabilities propagate from perception to policy and actuation.
• Architectural similarity, domain shift, and feedback dynamics critically influence vulnerability transferability and failure modes.

Conclusions:

• Adversarial threats in agentic AI are system-level risks requiring a shift from benchmark-centric evaluation.
• The proposed framework integrates control-theoretic reasoning and governance-aware defense design for agentic AI security.
• Findings have direct implications for safety-critical applications like autonomous mobility, healthcare imaging, and biometric security.