Skip to content
Projects

Yuxuan2003/Awesome-GUI-Agent-Security

View on GitHub

Yuxuan2003/Awesome-GUI-Agent-Security has 60 GitHub stars and 1 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-18 12:57 UTC. Last push 2026-11-02.

How these counts are measured

Measured star growth

in 24h
0
in 7 days
Collecting history
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

GUI / Computer-Use / Browser Agent 安全论文清单 —— 按攻防轴组织,每篇附中文简介

AI

README.md

Original

Awesome-GUI-Agent-Security

English | 简体中文

A curated list of papers on GUI / Computer-Use / Browser Agent security — organized by attack surface and defense layer, not by runtime environment.

Last Update Papers Time Range Link Check Awesome

This page is the index — one line per paper. Each section links to a page with a 3–5 sentence summary per paper.

Scope: papers whose primary subject is a GUI / computer-use / browser / mobile agent, with a security contribution. Not included: general LLM/agent security that only uses GUI agents as a testbed · agents for security work (pentest, CTF) · pure capability work.

Why organized by attack surface instead of environment?

Most GUI agent lists split papers by runtime environment (Web / Mobile / Desktop), which scatters a single attack class across sections: multi-step indirect injection lands under Desktop, efficiency backdoors under Mobile, pop-up attacks under both Web and Desktop. Answering "what visual-layer attacks exist?" means reading every section.

Here the primary axis is attack vector and defense intervention point. Runtime environment is a cross-cutting tag, used as a primary dimension only inside the benchmarks chapter.

Contents

Browse by environment: Web | Mobile | Desktop | Cross-env


0 Surveys & Threat Models

Surveys, SoKs, and mappings onto threat taxonomies such as OWASP ASI and MITRE ATLAS · Summaries →

  • WebMASLab — From Monoliths to Swarms: A Study of Attack Surface Evolution in the Transition to Multi-Agent Web Systems · 2026-07 · 🌐
  • SafeFlow — Semantic Information-Flow Control for Blocking Malicious Propagation in Multi-Agent Systems · 2026-07 · 🧩
  • Architecture-Lifecycle Framework — Securing Computer-Use Agents: A Unified Architecture-Lifecycle Framework for Deployment-Grounded Reliability · 2026-05 · 🧩
  • Mobile Agent Security Study — Measuring the Security of Mobile LLM Agents under Adversarial Prompts from Untrusted Third-Party Channels · 2025-10 · 📱
  • CUA Vuln SoK — A Systematization of Security Vulnerabilities in Computer Use Agents · 2025-07 · 🖥️🌐
  • JARVIS or Ultron? — A Survey on the Safety and Security Threats of Computer-Using Agents: JARVIS or Ultron? · 2025-05 · 🧩
  • Trustworthy GUI Survey — Towards Trustworthy GUI Agents: A Survey · 2025-03 · 🧩
  • BrowserART — Refusal-Trained LLMs Are Easily Jailbroken As Browser Agents · 2024-10 · 🌐

1 Attack Surfaces

Organized by attack vector and entry point, not by runtime environment

1.1 Indirect Prompt Injection

Injection carried by external content: web pages, documents, email · Summaries →

  • SIR — Self-improving Red-teaming for Compute Use Agents · 2026-08 · 🖥️🌐
  • StepJack — Benchmarking Computer-Use Agent Safety Against Multi-Step Indirect Prompt Injection · 2026-08 · 🖥️🧩
  • Invisible Ink — Invisible Ink Threats: Adversarial Goals Behind Legitimate Tasks in Computer-Use Agents · 2026-08 · 🖥️
  • ADI — Agent Data Injection Attacks are Realistic Threats to AI Agents · 2026-07 · 🌐🖥️
  • WebTrap — Stealthy Mid-Task Hijacking of Browser Agents During Navigation · 2026-05 · 🌐
  • ReadSecBench — You Told Me to Do It: Measuring Instructional Text-induced Private Data Leakage in LLM Agents · 2026-03 · 🖥️
  • Plan Injection — Context manipulation attacks : Web agents are susceptible to corrupted memory · 2025-06 · 🌐
  • The Hidden Dangers of Browsing AI Agents · 2025-05 · 🌐
  • Fine-Print Injection — The Obvious Invisible Threat: LLM-Powered GUI Agents' Vulnerability to Fine-Print Injections · 2025-04 · 🌐
  • SUDO — sudo rm -rf agentic_security · 2025-03 · 🖥️🧩
  • WIPI — A New Web Threat for LLM-Driven Web Agents · 2024-02 · 🌐

1.2 Visual-Layer Attacks

Adversarial patches, pop-up lures, typographic attacks, screenshot poisoning · Summaries →

  • WebMirage — Adversarial Images Hijack Web Agents from Visual Grounding to Browser Execution · 2026-10 · 🌐
  • UI Desynchronization — When Agents See Differently: Exposing UI Desynchronization Threats in Mobile Agents · 2026-09 · 📱
  • AgentHijack — Visual Patch Attacks on Multimodal Computer-Use Agents · 2026-09 · 🖥️🌐
  • Perception-Fusion Gap — Do GUI Agents Believe Their Eyes? Diagnosing State-Belief Reliance on Pixels versus Structure · 2026-07 · 🌐📱🖥️
  • MIRAGE — Stealthy Visual Prompt Injection for Vulnerability Detection in Web Agents · 2026-06 · 🌐
  • PriceBlind — Penny Wise, Pixel Foolish: Bypassing Price Constraints in Multimodal Agents via Visual Adversarial Perturbations · 2026-04 · 📱
  • PRAC — Preference Redirection via Attention Concentration: An Attack on Computer Use Agents · 2026-04 · 🖥️🌐
  • Semantic UI Injection — Are GUI Agents Focused Enough? Automated Distraction via Semantic-level UI Element Injection · 2026-04 · 🧩
  • Visual Confused Deputy — Exploiting and Defending Perception Failures in Computer-Using Agents · 2026-03 · 🖥️
  • Agent-Only Perceptual Injection — Invisible to Humans, Triggered by Agents: Stealthy Jailbreak Attacks on Mobile Vision-Language Agents · 2025-10 · 📱
  • CPS — Cross-Modal Content Optimization for Steering Web Agent Preferences · 2025-10 · 🌐
  • WebInject — Prompt Injection Attack to Web Agents · 2025-05 · 🌐
  • MIP — MIP against Agent: Malicious Image Patches Hijacking Multimodal OS Agents · 2025-03 · 🧩
  • ARE — Dissecting Adversarial Robustness of Multimodal LM Agents · 2024-06 · 🌐

1.3 Environmental Injection

UI element injection, accessibility tree, spoofed notifications, overlays · Summaries →

  • AnTrap — Are Android GUI Agents Robust Against Runtime Anomalies? AnTrap: Evaluating Agents in Dynamic Adversarial Environments · 2026-08 · 📱
  • Not an A11y — How Android Accessibility Exposes Mobile AI Agents to Indirect Prompt Injection · 2026-08 · 📱
  • MIRAGE (Mobile) — MIRAGE: Context-Aware Prompt Injection against Mobile GUI Agents via User-Generated Content · 2026-05 · 📱
  • WAAA — WAAA! Web Adversaries Against Web Agents · 2026-05 · 🌐
  • eTAMP — Poison Once, Exploit Forever: Environment-Injected Memory Poisoning Attacks on Web Agents · 2026-04 · 🌐
  • TrickyArena — Investigating the Impact of Dark Patterns on LLM-Based Web Agents · 2025-10 · 🌐
  • Dynamic EIA — Environmental Injection Attacks against GUI Agents in Realistic Dynamic Environments · 2025-09 · 🌐
  • A11y Tree IPI — Manipulating LLM Web Agents with Indirect Prompt Injection Attack via HTML Accessibility Tree · 2025-07 · 🌐
  • AdInject — Real-World Black-Box Attacks on Web Agents via Advertising Delivery · 2025-05 · 🌐
  • EVA — Evolving Semantic Adversaries for Red-Teaming GUI Agents Against Environmental Injection Attacks · 2025-05 · 🧩
  • EIA — Environmental Injection Attack on Generalist Web Agents for Privacy Leakage · 2024-09 · 🌐
  • Caution for the Environment: Multimodal LLM Agents are Susceptible to Environmental Distractions · 2024-08 · 🧩

1.4 Privilege Escalation & Permission Abuse

OS-level escalation, cross-app privilege abuse, permission-dialog manipulation, TOCTOU · Summaries →

  • Authority Envelope — The Agent Clicks Approve: Injected Page Text Makes Browser Agents Authorize Access the Task Never Named · 2026-10 · 🌐
  • Allow to Achieve — "Allow" to Achieve, Over-Privileged Inadvertently: The Unintended Cost of Task-Completion-Driven Pop-up Decisions in Mobile GUI Agents · 2026-08 · 📱
  • AI Sees — (A)I Sees What You Don't: Exploiting New Attack Surfaces in Third-Party Mobile Agents · 2026-07 · 📱
  • PUSV — Temporal UI State Inconsistency in Desktop GUI Agents: Formalizing and Defending Against TOCTOU Attacks on Computer-Use Agents · 2026-04 · 🖥️
  • Atomicity for Agents — Exposing, Exploiting, and Mitigating TOCTOU Vulnerabilities in Browser-Use Agents · 2026-02 · 🌐
  • Action Rebinding — Mind the Gap: Action Rebinding Attacks against Android GUI Agents · 2026-01 · 📱

1.5 Data Exfiltration & Privacy

Credential theft, PII leakage, contextual-integrity violations, oversharing · Summaries →

  • AgentTell — Behavioural Side-Channel Leakage in Browser-Use Agents · 2026-09 · 🌐
  • LoginTrap — Uncovering Task-Agnostic Phishing-Style Indirect Prompt Injection Attacks against LLM-based Web Agents · 2026-08 · 🌐
  • Capable but Careless — Do Computer-Use Agents Follow Contextual Integrity? · 2026-06 · 🖥️
  • Scammer4U — "I Strongly Suspect This Website Is a Scam": Benchmarking PII Leakage and Detection without Defense in Autonomous Web Agents · 2026-05 · 🌐
  • MyPhoneBench — Do Phone-Use Agents Respect Your Privacy? · 2026-04 · 📱
  • WebPII — Benchmarking Visual PII Detection for Computer-Use Agents · 2026-03 · 🌐🖥️
  • SPILLage — Agentic Oversharing on the Web · 2026-02 · 🌐

1.6 Backdoors & Poisoning

Grounding backdoors, efficiency backdoors, memory poisoning · Summaries →

  • SynChain — Inducing Computer-Use Agent Systems to Construct Their Own Attack Chains · 2026-08 · 🖥️
  • MemVenom — Triggered Poisoning of Multimodal Memories in Web Agents · 2026-06 · 🌐
  • AgentRAE — Remote Action Execution through Notification-based Visual Backdoors against Screenshots-based Mobile GUI Agents · 2026-03 · 📱
  • SlowBA — An Efficiency Backdoor Attack towards VLM-based GUI Agents · 2026-03 · 📱🧩
  • Skill Trust Framework — Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward · 2026-02 · 🧩
  • VisualTrap — A Stealthy Backdoor Attack on GUI Agents via Visual Grounding Manipulation · 2025-07 · 📱🖥️
  • VIBMA — Poison Once, Control Anywhere: Clean-Text Visual Backdoors in VLM-based Mobile Agents · 2025-06 · 📱
  • AgentGhost — Hidden Ghost Hand: Unveiling Backdoor Vulnerabilities in MLLM-Powered Mobile GUI Agents · 2025-05 · 📱

1.7 Unintended Harm from Benign Instructions

No adversary involved — harm arising from the agent's own behavior on normal tasks · Summaries →

  • Preflight — What Does "Now" Mean to Your Computer-Use Agent? Stale Screenshots, Page Loads, and Which Check to Run Before a Click · 2026-10 · 🌐
  • CAVEAT — Towards Robust Computer-Use Agents in Incentive-Misaligned Environments · 2026-09 · 🌐
  • Nudge Susceptibility — A Dual-Process Perspective on Nudge Susceptibility in LLM-Based GUI Agents · 2026-09 · 🌐
  • Alignment Is Local — A Paired Diagnostic for GUI Agents under User Persuasion · 2026-07 · 📱🧩
  • OTora — A Unified Red Teaming Framework for Reasoning-Level Denial-of-Service in LLM Agents · 2026-05 · 🌐🖥️
  • OS-BLIND — The Blind Spot of Agent Safety: How Benign User Instructions Expose Critical Vulnerabilities in Computer-Use Agents · 2026-04 · 🖥️
  • AutoElicit — When Benign Inputs Lead to Severe Harms: Eliciting Unsafe Unintended Behaviors of Computer-Use Agents · 2026-02 · 🖥️
  • The Behavioral Fabric of LLM-Powered GUI Agents: Human Values and Interaction Outcomes · 2026-01 · 🌐
  • AgentBait — When Bots Take the Bait: Exposing and Mitigating the Emerging Social Engineering Attack in Web Automation Agent · 2026-01 · 🌐
  • DECEPTICON — How Dark Patterns Manipulate Web Agents · 2025-12 · 🌐
  • BLIND-ACT — Just Do It!? Computer-Use Agents Exhibit Blind Goal-Directedness · 2025-10 · 🧩
  • Dark Patterns Meet GUI Agents — LLM Agent Susceptibility to Manipulative Interfaces and the Role of Human Oversight · 2025-09 · 🌐
  • Characterizing Unintended Consequences in Human-GUI Agent Collaboration for Web Browsing · 2025-05 · 🌐

2 Defense Layers

Organized by where the defense intervenes in the execution chain

2.1 Input Filtering & Sanitization

Filtering or masking untrusted content before it enters the model context · Summaries →

  • MPIDetector — Detecting Multi-modal Prompt Injection via LLM-based Reasoning · 2026-10 · 🧩
  • PrivacyGuard — Source-Linked Sanitization and Locally Resolvable PII References for Cloud GUI Agents · 2026-10 · 📱
  • FocusMem — Factorizing Content, Readout, and Trust in Latent GUI Memory · 2026-08 · 📱🖥️🌐
  • UCM — Untrusted Content Masking for Web Agents with Security Guarantees · 2026-07 · 🌐
  • IPI Probe Controls — When AUC 0.998 Is Not Enough: A Candidate Evaluation Protocol for Hidden-State Probes of Indirect Prompt Injection in Multimodal Computer-Use Agents · 2026-06 · 🧩
  • CAPED — Context-Aware Privacy Exposure Defense for Mobile GUI Agents · 2026-06 · 📱
  • MaskClaw — Edge-Side Personalized Privacy Arbitration for GUI Agents with Behavior-Driven Skill Evolution · 2026-05 · 📱🖥️
  • WARD — Adversarially Robust Defense of Web Agents Against Prompt Injections · 2026-05 · 🌐
  • SnapGuard — Lightweight Prompt Injection Detection for Screenshot-Based Web Agents · 2026-04 · 🌐
  • WebAgentGuard — A Reasoning-Driven Guard Model for Detecting Prompt Injection Attacks in Web Agents · 2026-04 · 🌐
  • Cognitive Firewall — The Cognitive Firewall: Securing Browser Based AI Agents Against Indirect Prompt Injection Via Hybrid Edge Cloud Defense · 2026-03 · 🌐
  • Available but Invisible — Anonymization-Enhanced Privacy Protection for Mobile GUI Agents: Available but Invisible · 2026-02 · 📱
  • WebSentinel — Detecting and Localizing Prompt Injection Attacks for Web Agents · 2026-02 · 🌐
  • Rennervate — Attention is All You Need to Defend Against Indirect Prompt Injection Attacks in LLMs · 2025-12 · 🌐
  • DualTAP — A Dual-Task Adversarial Protector for Mobile MLLM Agents · 2025-11 · 📱
  • CORE — Reducing UI Exposure in Mobile Agents via Collaboration Between Cloud and Local LLMs · 2025-10 · 📱
  • GEM — Gaussian Embedding Modeling for Out-of-Distribution Detection in GUI Agents · 2025-05 · 🧩

2.2 Pre-execution Risk Assessment

World-model prediction, action risk scoring · Summaries →

  • AdvSim2Real — Training Web Agents Against Adaptive Prompt Injection in a Web World Model · 2026-10 · 🌐
  • SCOPE — Beyond Task Completion: Training Capable and Safe Computer-Use Agents · 2026-08 · 🧩
  • SeerGuard — A Safety Framework for Mobile GUI Agents via World Model Prediction · 2026-07 · 📱
  • Argus — Uncertainty Quantification for Computer-Use Agents: A Benchmark across Vision-Language Models and GUI Grounding Datasets · 2026-06 · 🧩
  • DUDE — Don't Click That: Teaching Web Agents to Resist Deceptive Interfaces · 2026-05 · 🌐
  • DeAction — When Actions Go Off-Task: Detecting and Correcting Misaligned Actions in Computer-Use Agents · 2026-02 · 🖥️
  • SafePred — A Predictive Guardrail for Computer-Using Agents via World Models · 2026-02 · 🖥️
  • MirrorGuard — Toward Secure Computer-Use Agents via Simulation-to-Real Reasoning Correction · 2026-01 · 🖥️
  • PolicyGuard — Learning Efficient Guardrails for Compliance · 2025-10 · 🌐
  • WebGuard — Building a Generalizable Guardrail for Web Agents · 2025-07 · 🌐

2.3 Runtime Interception & Access Control

Information-flow tracking, OS-level policy enforcement, sandboxing · Summaries →

  • SPIDES — Information-Flow Control for Browser Agents · 2026-10 · 🌐
  • From Observations to Authorized Actions: Measuring Source Information in Android Agents · 2026-10 · 📱
  • LGWM — Right Screen, Wrong Transition: World Models as Verifiers for GUI Agents · 2026-10 · 📱🧩
  • ZoneClaw — Mitigating Persistent Memory Attacks by Establishing Memory-Zoning in OpenClaw-Style Computer-Use Agents · 2026-09 · 🖥️🧩
  • Veer — Before Acting, Change the State: Prospective State Intervention for Web Agents under Deceptive Interfaces · 2026-09 · 🌐
  • MATE — Policy-Aware Security Auditing for Mobile Agents via Synthesis-Driven Trajectory Learning · 2026-09 · 📱
  • HazardAuditor — From Executable Threats to Safer Computer-Use Agents · 2026-09 · 🖥️🌐
  • Key-Step Supervision — Monitoring Web Agents Without Internal Signals: Observable Trajectories and Key-Step Supervision · 2026-09 · 🌐
  • CURA — Certified Runtime Alarms for Computer-Use Agents · 2026-08 · 🖥️
  • CRATE-S — Automated Trajectory Evaluation for Mobile Agents via Step-Level Consequence Reasoning and Aggregation · 2026-08 · 📱
  • Prismata — Confining Cross-Site Prompt Injection in Web Agents · 2026-07 · 🌐
  • BraveGuard — From Open-World Threats to Safer Computer-Use Agents · 2026-05 · 🖥️🌐
  • TEE-Backed Isolation — Constraining Host-Level Abuse in Self-Hosted Computer-Use Agents via TEE-Backed Isolation · 2026-05 · 🖥️
  • ceLLMate — Sandboxing Browser AI Agents · 2025-12 · 🌐
  • OS-Sentinel — Towards Safety-Enhanced Mobile GUI Agents via Hybrid Validation in Realistic Workflows · 2025-10 · 📱
  • CSAgent — Secure and Efficient Access Control for Computer-Use Agents via Context Space · 2025-09 · 🖥️
  • AgentSentinel — An End-to-End and Real-Time Security Defense Framework for Computer-Use Agents · 2025-09 · 🖥️
  • CUA-SHADE-Arena — Reliable Weak-to-Strong Monitoring of LLM Agents · 2025-08 · 🖥️
  • HarmonyGuard — Toward Safety and Utility in Web Agents via Adaptive Policy Enhancement and Dual-Objective Optimization · 2025-08 · 🌐
  • GuardAgent — Safeguard LLM Agents by a Guard Agent via Knowledge-Enabled Reasoning · 2024-06 · 🌐

2.4 Human-in-the-Loop & Confirmation

Confirmation before critical actions, approval gates, interruptibility · Summaries →

  • TIPO — Mobile GUI Agent Privacy Personalization with Trajectory Induced Preference Optimization · 2026-04 · 📱
  • PrivWeb — Unobtrusive and Content-aware Privacy Protection For Web Agents · 2025-09 · 🌐
  • VeriOS — Query-Driven Proactive Human-Agent-GUI Interaction for Trustworthy OS Agents · 2025-09 · 📱🖥️
  • InquireMobile — Teaching VLM-based Mobile Agent to Request Human Assistance via Reinforcement Fine-Tuning · 2025-08 · 📱
  • VerificAgent — Domain-Specific Memory Verification for Scalable Oversight of Aligned Computer-Use Agents · 2025-06 · 🖥️
  • Toward a Human-Centered Evaluation Framework for Trustworthy LLM-Powered GUI Agents · 2025-04 · 🌐

2.5 Post-hoc Recovery & Rollback

Failure attribution, state rollback, repair after harm has occurred · Summaries →

  • SCOUT — Synergizing Reasoning and Tool-Use for Computer-Use Safety · 2026-09 · 🖥️🧩
  • Verified Repair — From Blind Edits to Verified Repair: Building Trustworthy User-Side LLM Agents for Web Accessibility · 2026-07 · 🌐
  • CUADebug — Diagnosing and Repairing Computer-Use Agent Failures · 2026-07 · 🖥️
  • Agent Fingerprinting — Whose Agent Are You? Multi-Layer Fingerprinting and Attribution of Autonomous Web Agents · 2026-06 · 🌐
  • What Did It Actually Do — "What Did It Actually Do?": Understanding Risk Awareness and Traceability for Computer-Use Agents · 2026-03 · 🖥️
  • KillBench — Can We Stop Malicious AI? KILLBENCH: A Benchmark for External AI Kill Switch Feasibility · 2025-09 · 🌐

2.6 Formal Guarantees & Verification

Defenses with provable guarantees: formal verification, control-flow integrity, conformal risk control · Summaries →

  • SA-CUA — Towards Safe Autonomy for Computer Use · 2026-10 · 🖥️🧩
  • Secure-CUA — Controlling Untrusted Influence in Computer-Use Agents · 2026-10 · 🧩
  • AOHP — An Open-Source OS-Level Agent Harness for Personalized, Efficient and Secure Interaction · 2026-06 · 📱
  • SkillHarness — Harnessing Safe Skills for Computer-Use Agents · 2026-06 · 🖥️
  • CORA — Conformal Risk-Controlled Agents for Safeguarded Mobile GUI Automation · 2026-04 · 📱
  • DMAST — Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal Attacks · 2026-03 · 🌐
  • Aura — Blind Gods and Broken Screens: Architecting a Secure, Intent-Centric Mobile Agent Operating System · 2026-02 · 📱
  • NOVA — CaMeLs Can Use Computers Too: System-level Security for Computer Use Agents · 2026-01 · 🖥️
  • Building Browser Agents — Architecture, Security, and Practical Solutions · 2025-11 · 🌐
  • LaSM — Layer-wise Scaling Mechanism for Defending Pop-up Attack on GUI Agents · 2025-07 · 🖥️🌐

3 Benchmarks & Datasets

The only chapter where runtime environment serves as a primary organizing dimension

3.1 Comprehensive & Cross-environment

Benchmarks spanning multiple environments or threat classes · Summaries →

  • Mind2Web-Injection — Beyond the Verdict: Evidence-Aligned Evaluation of Visual Prompt-Injection Guardrails · 2026-09 · 🌐
  • ADeptS-Bench — Measuring the Trustworthiness of Computer Use Agents Across Devices · 2026-08 · 🖥️📱
  • OSGuard — A Benchmark for Safety in Computer-Use Agents · 2026-06 · 🖥️🌐
  • CUA-HandCrafted — Domain-Conditioned Safety in Frontier Computer-Using Agents: A 793-Episode Browser Benchmark, a Coding-Domain Cross-Reference, and a Reproducibility Audit of Recent Red-Teaming · 2026-06 · 🖥️🌐
  • AgentHazard — A Benchmark for Evaluating Harmful Behavior in Computer-Use Agents · 2026-04 · 🖥️
  • GUIGuard-Bench — Toward a General Evaluation for Privacy-Preserving GUI Agents · 2026-01 · 📱🖥️
  • SusBench — An Online Benchmark for Evaluating Dark Pattern Susceptibility of Computer-Use Agents · 2025-10 · 🌐🖥️
  • CUAHarm — Measuring Harmfulness of Computer-Using Agents · 2025-07 · 🖥️
  • RiOSWorld — Benchmarking the Risk of Multimodal Computer-Use Agents · 2025-05 · 🖥️
  • RedTeamCUA — Realistic Adversarial Testing of Computer-Use Agents in Hybrid Web-OS Environments · 2025-05 · 🧩
  • DoomArena — A framework for Testing AI Agents Against Evolving Security Threats · 2025-04 · 🧩

3.2 Web Environment

Security evaluation for web / browser agents · Summaries →

  • BrowseSafe++ — Benchmarking Prompt Injection Detection Across Web Contexts, Languages, and Browser Views · 2026-10 · 🌐
  • Who Pays the Price — Who Pays the Price? Stakeholder-Centric Prompt Injection Benchmarking for Real-world Web Agents · 2026-06 · 🌐
  • WebDecept — Benchmarking Web Agent Safety under E-commerce Deceptive Interfaces · 2026-04 · 🌐
  • RiskWebWorld — A Realistic Interactive Benchmark for GUI Agents in E-commerce Risk Management · 2026-04 · 🌐
  • WebSP-Eval — Evaluating Web Agents on Website Security and Privacy Tasks · 2026-04 · 🌐
  • ClawTrap — A MITM-Based Red-Teaming Framework for Real-World OpenClaw Security Evaluation · 2026-03 · 🌐
  • MUZZLE — Adaptive Agentic Red-Teaming of Web Agents Against Indirect Prompt Injection Attacks · 2026-02 · 🌐
  • MalURLBench — A Benchmark Evaluating Agents' Vulnerabilities When Processing Web URLs · 2026-01 · 🌐
  • WebTrap Park — An Automated Platform for Systematic Security Evaluation of Web Agents · 2026-01 · 🌐
  • TRAP — It's a TRAP! Task-Redirecting Agent Persuasion Benchmark for Web Agents · 2025-12 · 🌐
  • BrowseSafe — Understanding and Preventing Prompt Injection Within AI Browser Agents · 2025-11 · 🌐
  • Genesis — Evolving Attack Strategies for LLM Web Agent Red-Teaming · 2025-10 · 🌐
  • SecureWebArena — A Holistic Security Evaluation Benchmark for LVLM-based Web Agents · 2025-10 · 🌐
  • WAInjectBench — Benchmarking Prompt Injection Detections for Web Agents · 2025-10 · 🌐
  • WAREX — Web Agent Reliability Evaluation on Existing Benchmarks · 2025-09 · 🌐
  • RISK — A Framework for GUI Agents in E-commerce Risk Management · 2025-09 · 🌐
  • WebRRSBench — Benchmarking MLLM-based Web Understanding: Reasoning, Robustness and Safety · 2025-09 · 🌐
  • WASP — Benchmarking Web Agent Security Against Prompt Injection Attacks · 2025-04 · 🌐
  • SafeArena — Evaluating the Safety of Autonomous Web Agents · 2025-03 · 🌐
  • AdvAgent — Controllable Blackbox Red-teaming on Web Agents · 2024-10 · 🌐
  • ST-WebAgentBench — A Benchmark for Evaluating Safety and Trustworthiness in Web Agents · 2024-10 · 🌐

3.3 Mobile Environment

Security evaluation for mobile / Android / iOS agents · Summaries →

  • MobileInject — Evaluating Reasoning-to-Action Security of MLLM-Based Mobile Agents under Cross-Modal Consistency Injection · 2026-10 · 📱
  • MobileJailBench — Red Queen Jailbreak Benchmarking for Mobile Agents · 2026-10 · 📱
  • PriMobiBench — Characterizing Visual Privacy Leakage in VLM-Driven Mobile GUI Agents · 2026-09 · 📱
  • MobileWorldSafety — Benchmarking GUI Agent Safety Against Environmental Injection Attacks in Android Apps · 2026-08 · 📱
  • GhostEI-Bench — Do Mobile Agents Resilience to Environmental Injection in Dynamic On-Device Environments? · 2025-10 · 📱
  • SAPA-Bench — Mind the Third Eye! Benchmarking Privacy Awareness in MLLM-powered Smartphone Agents · 2025-08 · 📱
  • MVISU-Bench — Benchmarking Mobile Agents for Real-World Tasks by Multi-App, Vague, Interactive, Single-App and Unethical Instructions · 2025-08 · 📱
  • AgentScan — From Assistants to Adversaries: Exploring the Security Risks of Mobile LLM Agents · 2025-05 · 📱
  • MobileSafetyBench — Evaluating Safety of Autonomous Agents in Mobile Device Control · 2024-10 · 📱
  • Systematic Categorization, Construction and Evaluation of New Attacks against Multi-modal Mobile GUI Agents · 2024-07 · 📱

3.4 Desktop & OS Environment

Security evaluation for desktop / OS-level computer-use agents · Summaries →

  • OS-SPEAR — A Toolkit for the Safety, Performance, Efficiency, and Robustness Analysis of OS Agents · 2026-04 · 🖥️
  • LPS-Bench — Benchmarking Safety Awareness of Computer-Use Agents in Long-Horizon Planning under Benign and Adversarial Scenarios · 2026-02 · 🖥️
  • OS-Harm — A Benchmark for Measuring Safety of Computer Use Agents · 2025-06 · 🖥️
  • VPI-Bench — Visual Prompt Injection Attacks for Computer-Use Agents · 2025-06 · 🖥️🌐

4 Commercial AI Browsers & Product Security

Primarily non-arXiv sources: vendor security advisories, CVEs, security blogs, disclosures. See docs/MAINTENANCE.md for how this chapter is tracked. · Summaries →


Contributing

Edit data/papers.yaml only — README.md, README.zh-CN.md, and everything under docs/ are generated by GitHub Actions. See CONTRIBUTING.md for inclusion criteria and entry format, and MAINTENANCE.md for the update workflow.

Related lists

This list focuses on the security of GUI/CUA agents. For adjacent areas:

  • General agent security (full OWASP ASI spectrum): LLMSecurity/awesome-agent-skills-security
  • Using agents for security work (red teaming / pentest): kagnlp/Awesome-Agentic-Security
  • Agent auditing and provenance: yzhao062/awesome-auditable-ai
  • GUI agent capability research: OSU-NLP-Group/GUI-Agents-Paper-List