acceptodds
Under review as a conference paper at ICLR 2027

DuelBench: A Benchmark for Closed-Loop Attack and Defense Between LLM Agents on Web Services

Abstract

Large language model (LLM) agents can exploit web vulnerabilities, but most benchmarks evaluate them against targets with fixed defenses. This leaves open how an attacker continues when a defender changes the rules during an attack. We introduce DuelBench, a benchmark for closed-loop attack and defense on live web services. It contains 44 scenarios across three domains and 11 weakness classes. At a controlled response interval, the defender reads traffic logs and updates blocking rules, while the attacker uses service feedback to choose its next action. We evaluate attack success, the cost of continuing under defense, and disruption to benign traffic. Experiments across attacker and defender pairings show that defense reduces success and increases blocked requests and identity use on multi-step tasks. Attacker rankings also change across defenders, while short attacks remain largely unaffected by the tested response schedule.

Primary area:
datasets and benchmarks
Keywords:
large language model agents, security benchmark, adversarial multi-agent, agent red-teaming

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_iRJBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents DuelBench, a novel benchmark designed to evaluate LLM agents in closed-loop web attack and defense interactive scenarios on live services.

Critical strengths

Unlike existing benchmarks that evaluate LLM agent exploits against static target defenses, DuelBench introduces live defender agents that dynamically inspect logs and update mitigation rules during an attack.

Critical weaknesses

The target service audit logs explicitly label incoming requests with attacker and benign tags. In real-world security monitoring, a defender's primary challenge is identifying adversarial traffic amidst normal user noise without prior ground-truth tags.

Questions for authors
  1. The proposed framework is suspected to be generated by AI. The author needs to declare whether the proposed framework uses AI graphics. The author only stated in the statement that AI is used to polish and write code, and I am not sure if this includes the use of AI for drawing frameworks.

  2. The results show that short-horizon attacks maintain near 100% ASR across all defense conditions. Does this structural delay mean that DuelBench is currently unable to evaluate defensive capabilities against single-step or short-horizon vulnerabilities?

  3. How do the authors justify the inclusion of these short scenarios in the overall benchmark average if the defensive agent has no opportunity to intervene?

  4. In realistic cybersecurity settings, a defender's primary challenge is detecting anomalies and attributing threats from noisy, unlabeled logs. How significantly does providing perfect ground-truth source tags artificially boost the performance of defender agents?

1 comment

Thank you for your detailed feedback on our manuscript. Below, we respond to each of your points.

The target service audit logs explicitly label incoming requests with attacker and benign tags. In real-world security monitoring, a defender's primary challenge is identifying adversarial traffic amidst normal user noise without prior ground-truth tags.

Source labels indicate which past requests came from the attacker, but they do not tell the defender which blocking rule will be effective. For example, blocking a shared page may stop the attack but also prevent normal users from accessing it. Blocking the attacker's IP keeps the page available, but the attacker may switch to another IP.

Therefore, the defender still needs to choose an effective response. Some studies [R1][R2][R3] also evaluate agents on fixing known vulnerabilities. They provide vulnerability details and test whether the agent can fix the issue.

To further address your concern, we test defense without source labels and discuss the results in our response to Q4 below.

[R1] Andy K. Zhang, et al., BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems. NeurIPS 2025.

[R2] Hwiwon Lee, et al., SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks. NeurIPS 2025.

[R3] Zheng Yu, et al., PatchAgent: A Practical Program Repair Agent Mimicking Human Expertise. USENIX Security 2025.

Q1. The proposed framework is suspected to be generated by AI. The author needs to declare whether the proposed framework uses AI graphics.

We used AI to generate the two robot icons in the framework figure. The other icons come from publicly available online resources. We created the framework's structure and layout manually in PowerPoint and exported it as a PDF.

We already selected “Yes, to aid or polish writing” under AI Assistance in the submission form. We considered the use of AI for these icons to fall within this category. In the revised version, we have clarified this use of AI in the AI Use Statement section.

Q2. The results show that short-horizon attacks maintain near 100% ASR across all defense conditions. Does this structural delay mean that DuelBench is currently unable to evaluate defensive capabilities against single-step or short-horizon vulnerabilities?

DuelBench can evaluate defense against short attacks that leave time to respond. A successful first-turn attack leaves no such chance under the schedule in Section 3.2. Its ASR therefore does not measure the defender's ability to stop it.

Besides, we tested a DeepSeek-V4-Pro attacker against a Qwen3.6-Flash defender with labels and k=1. We ran each of the four short-attack scenarios 10 times. All attacks succeeded.

ScenarioNo response opportunityAt least one response opportunity
Command injection37
JWT confusion010
Login SQL injection91
OAuth redirect64
Total1822

Therefore, the high ASR reflects both a lack of response time and unsuccessful defense. We have added this analysis to Appendix D.5 of the revised version.

Q3. How do the authors justify the inclusion of these short scenarios in the overall benchmark average if the defensive agent has no opportunity to intervene?

The averages used for our main comparisons already exclude short scenarios. Tables 3 and 5 report the multi-step mean, and Figure 3 averages the same five multi-step scenarios. Appendix D.3 defines this averaging rule. Short scenarios are reported separately to show the limits of response timing.

We have made this point more explicit in the captions of Tables 3 and 5 and Figure 3 in the revised version.

Q4. How significantly does providing perfect ground-truth source tags artificially boost the performance of defender agents?

We run an additional comparison with a DeepSeek-V4-Pro attacker and k=1. Each scenario has ten runs. We compare logs with and without labels, keeping all other settings the same.

DefenderMetricNo defense (%)With labels (%)Without labels (%)
DeepSeek-V4-ProOverall ASR64.555.556.4
DeepSeek-V4-ProMulti-step ASR58.040.044.0
DeepSeek-V4-ProFPR0.000.002.91
Qwen3.6-FlashOverall ASR64.555.560.9
Qwen3.6-FlashMulti-step ASR58.042.056.0
Qwen3.6-FlashFPR0.000.000.00

The effect of source labels varies across defenders. With labels, Qwen allows fewer attacks to succeed, especially on multi-step tasks. DeepSeek's ASR changes little, but it blocks fewer normal requests. Without labels, both defenders still reduce overall ASR compared with no defense.

We have added this comparison to Appendix E.2 of the revised version.

Reviewer_Z5qFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

DualBench is a cybersecurity benchmark where there is a "defender" that does some action after some number of turns of the attacker agent.

Critical strengths

Cybersecurity is a critical capability to measure with the recent improvements in AI agents. This is a timely benchmark.

Critical weaknesses

An AI agent benchmark must ensure that the tasks are valid and that the measurements correspond to what the reader expects. I do not have confidence that this benchmark does this.

A cybersecurity benchmark in particular must:

  1. Ensure that the grader only passes valid attacks
  2. Ensure that passes correspond to correct attacks.
  3. Ensure realism of the attack situation (and in this case defense situation). The paper does not contain enough details to determine if any of the three hold.
Questions for authors

N/A

3 comments

Thank you for your feedback. Below, we address each comment.

1. Ensure that the grader only passes valid attacks.

DuelBench uses explicit success conditions to check whether an attack is valid, which is an established practice [R1][R2][R3]. As described in Section 3.3, the service records whether the required objective is reached, and the grader checks these records.

For example, the private-note task requires access to another user's protected note. Reading a public note does not meet this goal. This makes the success criterion explicit and testable.

[R1] Zhun Wang, et al., CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale. ICLR 2026.

[R2] Yuxuan Zhu, et al., CVE-Bench: A Benchmark for AI Agents' Ability to Exploit Real-World Web Application Vulnerabilities. ICML 2025.

[R3] Hwiwon Lee, et al., SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks. NeurIPS 2025.

2. Ensure that passes correspond to correct attacks.

DuelBench does check whether successful attacks depend on the intended weakness. As described in Section 3.3, we do this through paired tests on vulnerable and fixed service versions. If the same attack succeeds on the vulnerable version but fails on the fixed version, this provides evidence that its success depends on the intended weakness. This validation principle is also used in CyberGym [R1], SEC-bench [R3], and BountyBench [R4].

We further extend these tests to more scenarios and an LLM attacker. Each pair uses the same task and environment seed.

AttackerTest casesSuccesses on vulnerable versionSuccesses on fixed version
Scripted44 scenarios, 3 seeds each1320
DeepSeek-V4-Pro11 scenarios, 10 seeds each710

All scripted pairs succeed only on the vulnerable version. Every LLM run on the fixed version fails. Pairs that fail in both versions are not counted as validated attacks. These tests provide evidence that the successful attacks depend on the service weakness.

[R4] Andy K. Zhang, et al., BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems. NeurIPS 2025.

3. Ensure realism of the attack and defense situation.

Following common practice in cyber agent evaluation [R5][R6][R7], DuelBench captures key mechanisms of real web attacks and defenses in a controlled environment.

As described in Section 3.1, our tasks implement known web weaknesses. For example, the IDOR task allows access to another user's private note because the service does not check its owner. Requests are executed against the service, and defense rules affect later requests, including normal traffic.

Compared with previous work, our contribution is to evaluate how an LLM attacker and defender respond to each other during the attack. In DuelBench, the defender reads logs and changes blocking rules. The attacker then decides how to continue. This lets us measure how defense decisions and response timing affect attack success and normal service use.

[R5] Xiaoxue Ren, et al., HackWorld: Evaluating Computer-Use Agents on Exploiting Web Application Vulnerabilities. ICLR 2026.

[R6] Zicheng Liu, et al., PACEbench: A Framework for Evaluating Practical AI Cyber-Exploitation Capabilities. ICLR 2026.

[R7] Elizabeth Bates, et al., Beyond Rewards in RL for Cyber Defence. ICML 2026.

CyberGym is a perfect example of a benchmark that has major flaw: https://www.proximal.so/blog/cybergym-cyberrl/

Since the authors follow prior work, which has major flaws, I can only assume this work also has major flaws.

I strongly recommend rejecting this paper.

Thank you for your response. It is worth noting that every benchmark has its own limitations. However, CyberGym was peer reviewed and selected as an oral paper at ICLR 2026. It has also drawn broad interest in the research community, with over 1k stars on GitHub. This suggests that CyberGym is a valuable paper for reference rather than just “a paper with major flaws”.

Due to the limited time for rebuttal, we have not been able to verify the results reported in the blog post you cited. However, we have already addressed some of CyberGym’s limitations in our paper. For example, compared to CyberGym’s crash-based grader, our grader uses service records to check whether an attack reached the goal specified for that task. This avoids counting an unrelated crash as a successful attack. We also measure how the attacker responds when the defender changes a blocking rule and whether that rule blocks normal users. With this design, DuelBench provides a controlled setting for evaluating web security agents against an active opponent.

We would appreciate it if you could reconsider your assessment based on our response.

Reviewer_M2WVOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

DuelBench introduces a closed-loop web attack-and-defense benchmark where defenders update blocking rules during ongoing attacks, enabling evaluation of attacker adaptation, resource use, and benign-traffic disruption.

Critical strengths

The problem is well motivated, and the closed-loop design fills a clear gap in existing cybersecurity benchmarks. The findings on defender-dependent attacker rankings and response timing are practically relevant. Joint reporting of attack success, resource use, and benign-traffic disruption provides a more informative evaluation than success-only metrics.

Critical weaknesses

Compared with PACEbench and BountyBench, DuelBench's distinctive contribution is closed-loop attacker-defender interaction. However, defenders receive labels distinguishing attack from benign traffic, so the evaluation primarily measures mitigation decisions under privileged observations. This limits the generalizability of its findings to realistic defensive environments.

Questions for authors

How would defender performance and the ASR–FPR trade-off change without attack/benign traffic labels?

1 comment

Thank you for your thoughtful feedback. Below, we respond to each of your concerns.

Q1. The evaluation primarily measures mitigation decisions under privileged observations. This limits the generalizability of its findings to realistic defensive environments.

We mainly study how an attacker and a defender respond to each other during an ongoing attack, not just the defender's performance. Source labels identify past attack requests, but they do not specify which blocking rule will work.

Therefore, a known threat still requires an effective response from the defender. For example, BountyBench [R1] provides details of a known vulnerability and evaluates whether an agent can fix it through code changes.

We also compare defense performance with and without source labels in our response to Q2 below. Both defenders still have lower overall ASR without labels than the no-defense baseline, while the effect of removing labels differs across defenders. In future work, we will consider more varied normal traffic and larger web services to better reflect realistic web environments.

[R1] Andy K. Zhang, et al., BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems. NeurIPS 2025.

Q2. How would defender performance and the ASR–FPR trade-off change without attack/benign traffic labels?

In the table below, we test how removing source labels affects ASR and FPR. We use a DeepSeek-V4-Pro attacker with k=1 and ten environment seeds for each of the 11 base scenarios. Both log settings use neutral IP and account IDs, with timestamps and attacker-turn counts hidden. All other settings are the same. FPR is the percentage of normal requests blocked.

DefenderMetricNo defense (%)With labels (%)Without labels (%)
DeepSeek-V4-ProOverall ASR64.555.556.4
DeepSeek-V4-ProMulti-step ASR58.040.044.0
DeepSeek-V4-ProFPR0.000.002.91
Qwen3.6-FlashOverall ASR64.555.560.9
Qwen3.6-FlashMulti-step ASR58.042.056.0
Qwen3.6-FlashFPR0.000.000.00

Both defenders still have lower overall ASR without labels than the no-defense baseline. DeepSeek shows little change in ASR but blocks some normal requests. Qwen allows more attacks to succeed, especially on multi-step tasks, while its FPR remains zero. We have added this comparison to Appendix E.2 of the revised version.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.