Robustness of Agentic AI Systems via Adversarially-Aligned Jacobian Regularization

ABSTRACT

This research investigates robustness issues in agentic AI systems where autonomous agents interact with dynamic environments and external tools. The authors introduce adversarially-aligned Jacobian regularization to stabilize agent behavior under adversarial perturbations. The approach reduces the risk of unpredictable policy changes caused by malicious inputs or environment shifts. Experimental results demonstrate improved stability and reliability in multi-agent decision scenarios.