Dissecting Adversarial Robustness of Multimodal LM Agents

As language models (LMs) are used to build autonomous agents in real environments, ensuring their adversarial robustness becomes a critical challenge. Unlike chatbots, agents are compound systems with multiple components, which existing LM safety evaluations do not adequately address. To bridge this gap, we manually create 200 targeted adversarial tasks and evaluation functions in a realistic threat model on top of VisualWebArena, a real environment for web-based agents. In order to systematically examine the robustness of various multimodal we agents, we propose the Agent Robustness Evaluation (ARE) framework. ARE views the agent as a graph showing the flow of intermediate outputs between components and decomposes robustness as the flow of adversarial information on the graph. First, we find that we can successfully break a range of the latest agents that use black-box frontier LLMs, including those that perform reflection and tree-search. With imperceptible perturbations to a single product image (less than 5% of total web page pixels), an attacker can hijack these agents to execute targeted adversarial goals with success rates up to 67%. We also use ARE to rigorously evaluate how the robustness changes as new components are added. We find that new components that typically improve benign performance can open up new vulnerabilities and harm robustness. An attacker can compromise the evaluator used by the reflexion agent and the value function of the tree search agent, which increases the attack success relatively by 15% and 20%. Our data and code for attacks, defenses, and evaluation are available at https://github.com/ChenWu98/agent-attack