The growing integration of artificial intelligence into critical infrastructure, finance, and security-sensitive environments introduces a new and complex attack surface that conventional cybersecurity approaches alone cannot adequately address. The proliferation of AI in smart cities, finance, and public services amplifies the need to examine security threats targeting both the models and their ecosystems. While AI systems offer capabilities such as real-time analytics and automated decision-making, they also introduce unique vulnerabilities.

In a paper presented at the IEEE Mediterranean Electrotechnical Conference, researchers examine a broad spectrum of AI-specific threats, including adversarial machine learning, AI-amplified social engineering, and AI-enhanced malware. The authors analyze inherent vulnerabilities in AI frameworks, covering data-related issues and model-related challenges. 

This synthesis aims to inform efforts to secure AI-enabled systems against evolving threats while balancing operational effectiveness with ethical obligations. Core security goals—confidentiality, integrity, availability, and resilience—remain central but must now be reinforced with AI-specific approaches. Emerging methods such as federated learning and homomorphic encryption aim to improve privacy and trust, underscoring the need to balance performance, security, and reliability in modern AI system design.

Key Threats

The authors identify several AI-specific threat categories:

  • Adversarial machine learning attacks, such as evasion and poisoning, where subtle input manipulations or corrupted training data undermine model reliability. 
  • AI-amplified social engineering erodes trust in digital communications, particularly deepfakes and highly personalized phishing. 
  • AI-enhanced malware uses adaptive and learning-based techniques to evade detection and dynamically exploit systems. 
  • Large Language Models introduce additional risks, including prompt injection, model inversion, jailbreaking, and supply-chain compromise via third-party models and plugins.  

Underlying these threats are structural vulnerabilities in AI systems. Data-related weaknesses include bias, fairness issues, and privacy leakage through inference or inversion attacks. Infrastructure vulnerabilities, ranging from insecure communications to exposed supply chains, expand the overall attack surface in complex AI deployments.  

Defense Strategies

Vulnerable IoT devices can be compromised to inject false data, leading to misinformed grid management decisions. To counter these risks, the authors argue for layered, adaptive defense mechanisms. Recommended approaches include advanced cryptography, blockchain-based immutable logging, self-learning AI defenses integrated with automated response platforms, deep-learning-based anomaly detection, and robust identity and access management with adaptive multi-factor authentication.  

The authors outline case studies in energy and finance that illustrate how AI attacks can directly impact physical systems and economic stability, emphasizing the need for domain-specific safeguards. Qualified analysis of AI security methods indicates that hybrid defense frameworks offer superior resilience. AI-based intrusion detection shows 10-20% higher accuracy than traditional systems. However, computational overhead remains a challenge. 

Future Trends and Research Directions

Looking ahead, the authors highlight that securing AI-enabled systems requires advancing research and policy priorities, including AI-blockchain integration, defensive uses of generative AI, and stronger ethical and regulatory frameworks. 

While AI enhances capabilities in critical systems, it also introduces new vulnerabilities, including adversarial attacks, AI-driven social engineering, and adaptive malware. Addressing these risks demands a multi-layered defense combining advanced cryptography, adaptive self-learning defenses, and robust authentication. Domain-specific protocols, especially in energy and finance, alongside ethical and legal governance, are essential to maintaining trust, privacy, and resilience in AI deployments.

Interested in learning more about Artificial Intelligence and Defense? We have thousands of articles related to these industries and more! Also learn more about our eBooks and eLearning collections on cutting edge technologies.

Interested in acquiring full-text access to this collection for your entire organization? Request a free demo and trial subscription for your organization.