Adversarial Attacks on AI Agents: Data Privacy Risks and Regulatory Implications
Hassan Abisoye Olugbile, Oghenemarho Anthony Karieren, Roqeeb Ayodeji Olaniyi, Esther Oluwapelumi Ayeku, Patience Odele · International Journal of Advances in Engineering and Management · 2025
The emergence of sophisticated security vulnerabilities is a consequence of the growing integration of autonomous AI agents in critical sectors, including finance and healthcare. The thorough overview of adversarial attacks especially aimed at these systems is given in this work We clarify the basic ideas of perturbations and adversarial instances, so showing how these small changes could fool artificial intelligence machines. Along with typical white-box and black-box tactics, the paper classifies several attack kinds including evasion, poisoning, model inversion, and newly agent-specific threats including quick injection and jailbreaking. The great data privacy concerns these attacks create—re-identification and the unanticipated use of artificial intelligencegenerated synthetic data as adversarial background knowledge—have a central focus. The article also looks at the changing regulatory scene including the EU AI Act and GDPR and describes important governance structures including NIST AI RMF. Acknowledging the ongoing "arms race" between adversaries and defenders, we underline the need of multifarious mitigating techniques including both sophisticated technical defenses and strong organizational regulations to build safe and trustworthy AI ecosystems.