Trajectory-Guided Framework Enhances AI Agent Risk Mitigation

As AI agents become increasingly integrated into organizational workflows, interacting with external information and digital tools, managing the risks associated with malicious or untrusted inputs is paramount. Current red-teaming methods often rely on fixed attack templates or focus solely on final outcomes, which limits insight into how vulnerabilities unfold across multi-step reasoning and tool use. This approach fails to capture the dynamic nature of agent execution risks. A new framework, TrajRed, addresses this by conceptualizing agent execution risk as a trajectory-level phenomenon. TrajRed is a trajectory-guided red-teaming framework designed to uncover vulnerabilities in agentic AI systems by analyzing their execution paths. Building on the insights gained from TrajRed, the researchers also developed TrajGuard, a runtime governance layer. TrajGuard utilizes the high-risk trajectories identified during red teaming to actively monitor ongoing workflows and intervene when necessary. Experiments conducted on AgentDojo across four organizational task suites demonstrated TrajRed's effectiveness, identifying substantially stronger vulnerabilities than traditional fixed-template and automatic red-team baselines. Furthermore, TrajGuard proved highly successful in mitigating these risks, reducing attack success across all evaluated methods to near zero while maintaining the utility of benign tasks. These results underscore the critical importance of understanding and governing agent execution trajectories for robust AI deployments in organizational settings.