Threats and vulnerabilities in artificial intelligence and agentic AI models

Executive Summary

This paper explores threats and vulnerabilities in artificial intelligence (AI) and agentic AI models, extending the threat model to autonomy and closed-loop decision-making. It highlights that no single defense mechanism provides robustness across all layers of agentic AI systems, emphasizing the need for multi-layered strategies.

Why It Matters

Understanding the vulnerabilities of AI and agentic AI systems is crucial for national security, as their exploitation can lead to cyber-warfare, espionage, and strategic destabilization.

Key Takeaways

• Adversarial exploitation of AI-enabled defense platforms poses significant geopolitical risks, including cyber-warfare, espionage, and strategic destabilization.
  Source evidence · PDF page 2

  At the geopolitical level, adversarial exploitation of AI-enabled defence platforms raises the spectre of cyber-warfare, espionage, and strategic destabilisation.

• Current research on agentic AI adversarial studies is in an early exploratory phase, with many higher-order attacks remaining underexplored empirically.
  Source evidence · PDF page 27

  The remaining author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

• No single defense mechanism is universally effective against AI adversarial attacks; multi-layered strategies combining various techniques are necessary for measurable resilience.
  Source evidence · PDF page 26

  No single defensive mechanism exhibits universal efficacy. Defensive distillation, feature squeezing, and gradient masking all deteriorate under adaptive evaluation; only multi-layered strategies that combine adversarial training, randomised smoothing, differential privacy, and interface hardening provide measurable resilience, and even these solutions leave a gap of more than 50 percentage points bet

Strategic Relevance

The document directly addresses the strategic implications of AI vulnerabilities for national defense and geopolitical stability, making it highly relevant for understanding future conflict domains and security challenges.