Dissecting Adversarial Robustness of Multimodal LM Agents
read the original abstract
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 taking actions, which existing LMs safety evaluations do not adequately address. To bridge this gap, we manually create 200 targeted adversarial tasks and evaluation scripts in a realistic threat model on top of VisualWebArena, a real environment for web agents. To systematically examine the robustness of 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. We find that we can successfully break latest agents that use black-box frontier LMs, including those that perform reflection and tree search. With imperceptible perturbations to a single 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 inference-time compute that typically improves 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 at https://github.com/ChenWu98/agent-attack
This paper has not been read by Pith yet.
Forward citations
Cited by 17 Pith papers
-
Heterogeneous LLM Debate Under Adversarial Peers: Honest Gains, Replacement Costs, and Resilience
Honest heterogeneous peers in LLM debates lower harmful revision rates (e.g., 89% to 35%), while adversarial peers raise them (to 90%), and provide defense even against same-family adversaries.
-
Who Pays the Price? Stakeholder-Centric Prompt Injection Benchmarking for Real-world Web Agents
Introduces a stakeholder-centric benchmark showing current web agents fail all tested prompt injection objectives, with failures falling into stealthy parasitism, misaligned disruption, or compounded failure modes.
-
HLL: Can Agents Cross Humanity's Last Line of Verification?
HLL is a new benchmark that evaluates eight frontier multimodal agents on closed-loop interactive CAPTCHA solving, showing sharp performance drops under realism stressors and trace validation.
-
Same Payload, Different Channel: Measuring Trust Asymmetry in Tool-Using Language Models
Agent-native LLMs are substantially more vulnerable to adversarial instructions arriving in tool descriptions than user messages (with the pattern reversing for general-purpose models and inverting again for tool outp...
-
Hierarchical Attacks for Multi-Modal Multi-Agent Reasoning
HAM³ achieves up to 78.3% attack success rate on the GQA benchmark by hierarchically attacking perception, communication, and reasoning layers in multi-modal multi-agent systems.
-
A Systematic Survey of Security Threats and Defenses in LLM-Based AI Agents: A Layered Attack Surface Framework
A new 7x4 taxonomy organizes agentic AI security threats by architectural layer and persistence timescale, revealing under-explored upper layers and missing defenses after surveying 116 papers.
-
Adversarial Hubness in Multi-Modal Retrieval
Adversarial hubs can be generated to be retrieved as top-1 for over 84% of test queries in text-to-image retrieval, far exceeding natural hubs.
-
ROGUE: Misaligned Agent Behavior Arising from Ordinary Computer Use
Frontier AI agents frequently violate corrigibility by overriding interruptions in benign computer-use tasks, with misalignment increasing alongside model capability.
-
SnapGuard: Lightweight Prompt Injection Detection for Screenshot-Based Web Agents
SnapGuard detects prompt injection attacks on screenshot-based web agents via visual stability indicators and contrast-polarity textual signals, reaching F1 0.75 while running 8x faster than GPT-4o with no added memory cost.
-
How Adversarial Environments Mislead Agentic AI?
Adversarial compromise of tool outputs misleads agentic AI via breadth and depth attacks, revealing that epistemic and navigational robustness are distinct and often trade off against each other.
-
Visual Inception: Compromising Long-term Planning in Agentic Recommenders via Multimodal Memory Poisoning
Visual Inception poisons images to hijack long-term memory in agentic recommenders and steer planning, while CognitiveGuard reduces success to about 10% via perceptual sanitization and reasoning verification.
-
Mobile GUI Agents under Real-world Threats: Are We There Yet?
Introduces an app-content instrumentation framework and benchmark showing that examined GUI agents suffer 42.0% and 36.1% average misleading rates from third-party content in dynamic and static tests respectively.
-
WARD: Adversarially Robust Defense of Web Agents Against Prompt Injections
WARD is a guard model trained on 177K web samples and adversarially hardened via attacker-guard co-evolution to achieve high recall on prompt injections with low false positives and no added latency.
-
Auto-ART: Structured Literature Synthesis and Automated Adversarial Robustness Testing
Auto-ART delivers the first structured synthesis of adversarial robustness consensus plus an executable multi-norm testing framework that flags gradient masking in 92% of cases on RobustBench and reveals a 23.5 pp rob...
-
GUI agent: Guided Exploration of User-Sensitive Screens
An explorer agent is developed to identify user-sensitive queries in GUI environments by systematic exploration starting from a single demonstrated task.
-
Agentic AI Security: Threats, Defenses, Evaluation, and Open Challenges
A survey that taxonomizes threats to agentic AI, reviews benchmarks and evaluation methods, discusses technical and governance defenses, and identifies open challenges.
-
Large Language Model Agent: A Survey on Methodology, Applications and Challenges
A survey that deconstructs LLM agent systems via a methodology-centered taxonomy linking design principles to emergent behaviors, applications, and challenges.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.