HackWorld: Evaluating Computer-Use Agents on Exploiting Web Application Vulnerabilities
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: We're looking at the paper: "HackWorld: Evaluating Computer-Use Agents on Exploiting Web Application Vulnerabilities".
Elias: Here's the gist of the paper. Web applications are prime targets for cyberattacks due to their role as entry points to vital services and sensitive data repositories HackWorld, the first evaluation framework for systematically assessing computer-use agents’ capabilities in exploiting web application vulnerabilities through visual interaction The evaluation of state-of-the-art CUAs reveals concerning patterns: CUAs achieve exploitation rates below 12% yet frequently show poor cybersecurity awareness during attempts This study introduces HackWorld, the first framework for systematically evaluating computer-use agents’ capabilities in exploiting web application vulnerabilities.
Nadia: Framework Introduction HackWorld is designed to systematically assess computer-use agents’ capabilities in exploiting web application vulnerabilities through visual interaction Unlike existing benchmarks using sanitized environments, HackWorld exposes CUAs to 36 curated applications spanning 11 frameworks and 7 languages, containing realistic vulnerabilities including injection flaws, authentication bypasses, and unsafe input handling The framework directly evaluates CUAs’ ability to discover and exploit these vulnerabilities using Capture-the-Flag (CTF) methodology while navigating complex web interfaces.
Elias: Evaluation Environment and Methodology HackWorld operates within a Kali Linux environment, providing industry-standard security tools used by cybersecurity practitioners The evaluation follows a structured pipeline: (1) Task Assignment, (2) Environment Perception, (3) Tool Selection and Execution, (4) Action Execution, and (5) Progress Monitoring.
Nadia: The cornerstone of HackWorld lies in its integration with Kali Linux’s extensive security toolkit This toolset enables systematic measurement of whether agents can select appropriate tools for specific contexts, interpret tool outputs accurately, and orchestrate multiple tools into coherent workflows.
Title and authors: Nadia: The title itself sets a very practical stage, focusing directly on "Exploiting Web Application Vulnerabilities." It’s not just about browsing; it’s about the actual attack part. And you see they're looking at computer-use agents specifically, which means they're testing how these things interact with visual interfaces.
Elias: I agree. The authors are listing several people from different universities and institutions, which tells you this is a collaborative effort pulling in security researchers and AI folks together to get a holistic view of the problem.
Priya: From my side, I wonder if focusing on exploitation rates gives you the full picture of what's actually going wrong with these agents. Are we measuring success enough?
Nadia: That’s a fair point, Priya. The paper points out that even with this setup, the agents often show security-oblivious behavior despite having technical capability to interact with vulnerable systems, which is something I think is really important to see in practice.
The paper's summary: Elias: The core of what they found, according to the summary, is that these state-of-the-art models can only achieve exploitation rates below twelve percent in their tests. That’s a pretty low number if you think about how much work it takes to find a real bug.
Nadia: Twelve percent is concerning because it suggests that while they can navigate the interface, they often lack the necessary cybersecurity reasoning to actually discover and exploit those flaws effectively. The paper emphasizes that agents frequently show poor cybersecurity awareness during their attempts, which is a big red flag for security applications.
Priya: So, it’s not just about failing to find the bug; it’s about them acting in a way that suggests they don't understand the security context of what they are doing, right? What kind of behavior did they observe that made you think that?
Elias: Well, the summary points out that this limitation comes down to fundamental issues in reasoning and planning rather than just being bad at seeing things on the screen. It highlights a gap between their ability to interact with a system and their actual cybersecurity knowledge.
The paper's improvements: Nadia: Now, the paper doesn't just stop at finding these limitations; they actually propose several ways to improve this evaluation framework itself, trying to fix the way we test these agents. They suggest making the environment more rigorous.
Elias: I think their main suggestions lean heavily toward forcing better tool usage and stricter verification of actions. They want to move beyond just seeing if an agent can click a button and start demanding that it analyzes its own outputs before taking the next step.
Priya: That makes sense from a measurement standpoint. If we only measure success, we miss the messy process of how they fail, so forcing them to show their work—the analysis of output—is crucial for getting any real data on their capabilities.
Nadia: Exactly. They want to see if we can enforce better decision-making loops instead of just looking at a final outcome. This moves the focus from simple task completion to how the agent reasons through complex security challenges step by step.
Conclusion: Elias: So, wrapping up, it seems the main message is that while agents can interact visually with web apps, they are currently lacking the strategic reasoning and planning needed for effective vulnerability discovery and exploitation. The paper points out that their performance is bottlenecked by these higher-level cognitive skills.
Nadia: And this means for us in security research, we need to focus less on just making the interface look good for an agent, and more on building systems that reward better decision-making in these complex workflows. We need agents that actually think like security professionals.
Priya: I'd add that the authors highlight a real gap between human-oriented tools and what these agents need; they suggest adopting Agent experience principles to target those specific failures in how the agent is designed to interact with things.
Elias: It sets a clear direction for future work: we should prioritize enhancing multi-step planning and decision-making capabilities over just trying to get better visual perception, because that seems to be where the gains are going to come from.
Nadia: So, this "HackWorld" framework provides a solid foundation for figuring out what skills are missing in modern AI agents when it comes to real penetration testing. It shows us exactly where the current limitations lie before we can build something more capable.
Zhejiang University · University of New South Wales University of New South Wales Institute National University of Singapore National University of Singapore CSIRO Australian National University Monash University
cs.CR, cs.CL
Submitted: 2025-10-14
Updated: 2026-11-13
Code: https://github.com/GUI-Agent/HackWorld
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 98/100
The gist: Web applications are prime targets for cyberattacks due to their role as entry points to vital services and sensitive data repositories HackWorld…
Key concepts
- HackWorld
- A new evaluation framework designed to systematically assess computer-use agents' ability to find and exploit web application flaws by interacting with a website visually. It exposes agents to 36 realistic applications containing common vulnerabilities like injection flaws.
- Capture-the-Flag (CTF) Methodology
- The method used in HackWorld where the agent must navigate a complex web interface to discover hidden flags. This requires the agent to use security tools and follow a structured process, mimicking real penetration testing scenarios.
- Agent eXperience (AX) Principles
- A proposed set of design principles for security tools that make them better for AI agents. These include ensuring tool outputs are machine-readable (like JSON) and clearly state errors, which helps agents reason more effectively and avoid simple failures.
Terminology
Summary
Web applications are prime targets for cyberattacks due to their role as entry points to vital services and sensitive data repositories HackWorld, the first evaluation framework for systematically assessing computer-use agents’ capabilities in exploiting web application vulnerabilities through visual interaction The evaluation of state-of-the-art CUAs reveals concerning patterns: CUAs achieve exploitation rates below 12% yet frequently show poor cybersecurity awareness during attempts This study introduces HackWorld, the first framework for systematically evaluating computer-use agents’ capabilities in exploiting web application vulnerabilities. p. 0 We introduce HackWorld, the first evaluation framework
for systematically assessing computer-use agents’ capabilities in exploiting web
application vulnerabilities through visual interaction. Unlike existing benchmarks
using sanitized environments, HackWorld exposes CUAs to 36 curated applications spanning 11 frameworks and 7 languages, containing realistic vulnerabilities
including injection flaws, authentication bypasses, and unsafe input handling.Where the paper says it
Framework Introduction
HackWorld is designed to systematically assess computer-use agents’ capabilities in exploiting web application vulnerabilities through visual interaction Unlike existing benchmarks using sanitized environments, HackWorld exposes CUAs to 36 curated applications spanning 11 frameworks and 7 languages, containing realistic vulnerabilities including injection flaws, authentication bypasses, and unsafe input handling The framework directly evaluates CUAs’ ability to discover and exploit these vulnerabilities using Capture-the-Flag (CTF) methodology while navigating complex web interfaces.
Evaluation Environment and Methodology
HackWorld operates within a Kali Linux environment, providing industry-standard security tools used by cybersecurity practitioners The evaluation follows a structured pipeline: (1) Task Assignment, (2) Environment Perception, (3) Tool Selection and Execution, (4) Action Execution, and (5) Progress Monitoring. The cornerstone of HackWorld lies in its integration with Kali Linux’s extensive security toolkit This toolset enables systematic measurement of whether agents can select appropriate tools for specific contexts, interpret tool outputs accurately, and orchestrate multiple tools into coherent workflows. p. 3 This toolset enables systematic measurement of whether agents
can select appropriate tools for specific contexts, interpret tool outputs accurately, and orchestrate
multiple tools into coherent workflows. Evaluation and Logging Infrastructure Throughout the evaluation process, HackWorld maintains
comprehensive logging and monitoring systems.Where the paper says it
Experimental Setup and Results
The evaluation employs two types of agents, including four topperforming proprietary models and two open-source GUI action models Observation spaces include Screenshot, Screenshot + a11ytree, and Set-of-Marks. Overall performance evaluation shows that Claude-3. p. 5 00%
Table 2 shows the overall performance of the five evaluated computer-use agents (i. , Claude-3. From a quantitative perspective, the results reveal substantial variation in performance. 7-Sonnet achieves the highest average success rate (10. 7-Sonnet achieves the highest average success rate (10. 18% across all observation spaces),
which is nearly double that of Claude-4-Opus (4. 18% across all observation spaces),
which is nearly double that of Claude-4-Opus (4. 63%) and over three times that of Claude-3. , Claude-3. 5-Sonnet (2024), Claude-3.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Key Findings on Agent Limitations
The analysis reveals several systematic failure patterns, including ineffective tool selection and output parsing, poor failure recovery and plan repair, and gaps in directory and source enumeration The ability of agents to control the computer is not the main bottleneck; rather, the upper performance limit for cybersecurity tasks is primarily limited by reasoning, planning, and tool orchestration capabilities rather than perceptual input. Efficient and selective tool usage is more informative than sheer frequency of tool calls, emphasizing that strategic tool integration is critical. p. 7 Key Insights. Overall, the analysis highlights three main observations: (1) efficient and selective tool
usage is more informative than sheer frequency of tool calls, emphasizing the importance of strategic
tool integration; (2) once basic perceptual fidelity is ensured, additional structuring of the observation
space (e.Where the paper says it
Implications for Future Work
The findings underscore a significant need for improvement in CUA cybersecurity skills and outline key research directions Future work should prioritize strategic reasoning and decision-making over better perception, as enhancing multi-step planning will yield greater gains. The failures in the evaluation reflect a mismatch between human-oriented tool UX and agent-oriented needs, suggesting that adopting Agent eXperience (AX) principles could directly target these failure modes. p. 9 Implications for Tool/Interface Design. The failures in our evaluation reflect a mismatch between
human-oriented tool UX and agent-oriented needs. These findings underscore a significant need for improvement in CUA
cybersecurity skills and outline key research directions. HackWorld thus establishes a vital foundation
for developing autonomous agents with advanced penetration testing capabilities.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Move beyond task completion metrics to measure reasoning and strategic decision-making by evaluating how agents
actively explore the environment, gather relevant information, and iteratively test their hypotheses.
This focuses onreasoning over feedback
rather than just perceptual fidelity as the primary bottleneck. -
Implement a system that enforces efficient and selective tool usage by requiring agents to analyze prior outputs before invoking tools, addressing the failure mode where agents
frequently executed multiple steps to launch different or identical tools, without adequately analyzing prior outputs.
-
Design a feedback mechanism that mandates
closed-loop observation
where agents must read and parse tool output before proceeding with any action, mitigating failures likeLack of closed-loop observation: All tools (browser, Burp, DirBuster, Nikto) simply 'open/click/start,' without reading and parsing the output.
-
Develop a standardized interface for security tools that enforces
Agent eXperience (AX) principles,
ensuring that outputs aremachine-readable
(JSON/JSONL) and includeexplicit state and error codes
to facilitate step-by-step reasoning loops. -
Enhance agent capabilities in authentication by requiring systematic attempts to establish or maintain sessions, including trying standard bypasses like
weak creds, SQLi login, password reset, JWT tamper,
rather than failing when sessions are not maintained.
Abstract
Web applications are prime targets for cyberattacks as gateways to critical services and sensitive data. Traditional penetration testing is costly and expertise-intensive, making it difficult to scale with the growing web ecosystem. While language model agents show promise in cybersecurity, modern web applications demand visual understanding, dynamic content handling, and multi-step interactions that only computer-use agents (CUAs) can perform. Yet, their ability to discover and exploit vulnerabilities through graphical interfaces remains largely unexplored. We present HackWorld, the first framework for systematically evaluating CUAs' capabilities to exploit web application vulnerabilities via visual interaction. Unlike sanitized benchmarks, HackWorld includes 36 real-world applications across 11 frameworks and 7 languages, featuring realistic flaws such as injection vulnerabilities, authentication bypasses, and unsafe input handling. Using a Capture-the-Flag (CTF) setup, it tests CUAs' capacity to identify and exploit these weaknesses while navigating complex web interfaces. Evaluation of state-of-the-art CUAs reveals concerning trends: exploitation rates below 12% and low cybersecurity awareness. CUAs often fail at multi-step attack planning and misuse security tools. These results expose the current limitations of CUAs in web security contexts and highlight opportunities for developing more security-aware agents capable of effective vulnerability detection and exploitation.
Sources
- EnIGMA: Interactive Tools Substantially Assist LM Agents in Finding Security Vulnerabilities
- Qwen2.5-VL Technical Report
- Purple Llama CyberSecEval: A Secure Coding Benchmark for Language Models
- Windows Agent Arena: Evaluating Multi-Modal OS Agents at Scale
- SeeClick: Harnessing GUI Grounding for Advanced Visual GUI Agents
- WASP: Benchmarking Web Agent Security Against Prompt Injection Attacks
- Frontier AI's Impact on the Cybersecurity Landscape
- A Systematization of Security Vulnerabilities in Computer Use Agents
- Scaling Laws for Neural Language Models
- SecQA: A Concise Question-Answering Dataset for Evaluating Large Language Models in Computer Security
- CAI: An Open, Bug Bounty-Ready Cybersecurity AI
- The Hidden Dangers of Browsing AI Agents
- UI-TARS: Pioneering Automated GUI Interaction with Native Agents
- AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents
- A Framework for Evaluating Emerging Cyberattack Capabilities of AI
- Learn-by-interact: A Data-Centric Framework for Self-Adaptive Agents in Realistic Environments
- OS-Genesis: Automating GUI Agent Trajectory Construction via Reverse Task Synthesis
- UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning
- OS-Copilot: Towards Generalist Computer Agents with Self-Improvement
- OS-ATLAS: A Foundation Action Model for Generalist GUI Agents
Related papers
- From Strategy Search to Policy Learning: MCTS-Guided Adversarial Co-Evolution for LLM Safety Alignment
- Privacy Leakage from Recurrent State in LLM Inference
- Incrementally Verifiable Oracle Aided Computation
- Privilege and confidentiality in generative AI workflows
- BadMoE: Backdooring Mixture-of-Experts LLMs via Routing-Aware Triggers and Task-Underutilized Experts
- CoVault: Secure, Scalable Analytics of Personal Data