FlagThis

Abstract

This talk addresses the non-deterministic risks associated with AI agents, specifically "Goal Hijacking" (via prompt injection) and "Rogue Agent" behavior (autonomous errors). The speaker proposes a runtime security architecture that functions as a "firewall" between an agent's reasoning process and its tool execution. By utilizing an ensemble of specialized, fine-tuned Small Language Models (SLMs) and a self-improving Retrieval-Augmented Generation (RAG) memory system, the proposed solution enables real-time, contextual security gating with low latency and reduced false positives.