Adversarial Fine‐Tuning and Model Reprogramming

This chapter exposes one of the most insidious threats in modern artificial intelligence (AI) systems: adversarial fine-tuning and model reprogramming. It explores how attackers exploit standard fine-tuning processes to subtly alter generative model behavior, embedding malicious, biased, or unauthorized outputs that activate only under specific conditions. Through targeted manipulation of training data, gradient updates, and prompt structures, adversaries can reshape the logic of a model without affecting its performance on traditional benchmarks—making detection both difficult and critical. Readers will gain a clear understanding of how semantic perturbations, prompt conditioning, and gradient-aligned Trojans allow covert behaviors to be embedded deep within transformer models. The chapter outlines how adversaries test and optimize these behaviors for persistence, stealth, and cross-model transferability, even across different architectures and deployment contexts. These techniques are not fringe cases—they represent a growing class of scalable, low-cost threats that bypass traditional security assumptions. The chapter also equips readers with a defender's mindset, offering strategic guidance for identifying behavioral drift, verifying model provenance, and stress-testing prompt responses. By framing reprogramming as a lifecycle risk—rather than an isolated event—it reinforces the need for continuous validation across the model development pipeline. This is essential reading for AI security professionals, risk managers, and engineers tasked with deploying safe and trustworthy generative systems in real-world environments.