acceptodds
Under review as a conference paper at ICLR 2027

ArenaWeaver: Automatic Generation of Environments and Attacks for Prompt Injection Evaluation at Scale

Abstract

As LLM agents are increasingly deployed in real-world applications, adversarial attacks such as indirect prompt injection pose significant security risks by manipulating agent behavior and potentially leading to severe consequences. Evaluating agents against these attacks requires controllable environments that reproduce how attacks unfold in realistic workflows, but manually constructing such environments is costly and hard to scale. We introduce ArenaWeaver, a framework that automatically simulates well-known applications from public documentation and supports flexible composition of these applications into interactive worlds for agent security evaluation, with browser and Model Context Protocol (MCP) access for each application. The suite spans 40 services, including simulations of popular email, messaging, payment, and social platforms, with 1,263 web-UI routes and 1,179 MCP tools. In addition, to accelerate red teaming tests in long-horizon tasks, we propose Counterfactual Forking Search, an adaptive tree search algorithm that leverages recorded interactions and agent feedback to select where to fork and how to deliver or revise injections. Evaluation of 25 frontier models on 660 curated cases reveals systematic differences in robustness. We find that successful attacks often remain effective across interfaces on compatible tasks, with higher overall transfer from browser to MCP (72.5%) than in reverse (60.4%). Across all evaluated models, attacks had higher success rates when they corrupted an action the user had requested than when they pursued an unrelated goal; in some cases, agents treated fabricated claims as evidence that an action met the user’s stated conditions. Overall, ArenaWeaver lays a foundation for co-evolution: stronger agents could build richer agent worlds and uncover more failure patterns to help improve agents.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
prompt injection, llm agents, agent security, automated environment generation, automated attack generation, adversarial evaluation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bW3POpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

ArenaWeaver automatically reconstructs applications for prompt injection evaluation, with browser and MCP interfaces sharing persistent state. It provides 40 applications and introduces Counterfactual Forking Search (CFS), which reuses recorded agent histories to explore injection locations and content. The paper evaluates 25 models on 660 curated cases and studies cross-model transfer, cross-interface transfer, and failures where fabricated evidence corrupts an otherwise authorized action.

Critical strengths
  1. The environment design enables useful security experiments that are difficult with simpler tool-response benchmarks. Browser and MCP access share actual application state, and attacks enter through separate application-level delivery tools. This makes it possible to trace outside content through an agent’s actions to persistent consequences, and to compare interfaces while preserving the task and intervention. The breadth of applications and composed workflows makes this a substantial infrastructure contribution.
  2. The evaluation offers useful findings with supporting controls and concrete evidence. Fixed attacks are evaluated from the task’s beginning rather than only from the searched continuation, and separate clean runs help distinguish capability from robustness. I particularly appreciated the finding that an agent can reject explicit outsider instructions yet accept fabricated evidence about whether a requested action is authorized (§4.2, App. H.8). The shared-task comparison and narrower delivery analyses also strengthen the reported robustness differences.
Critical weaknesses
  1. The evidence for CFS’s advantage as a search method is incomplete. Figure 3a compares shorter branch continuations with full benign runs. This demonstrates reduced victim execution per continuation, but does not establish better attack discovery under an equal total budget. A comparison against full-restart attack search, including attacker and setup costs, would better substantiate this contribution. This limits my assessment of CFS more than of the environment suite.
  2. Aggregate outcome validation still relies heavily on model judges. The blinded review supporting 595/628 successful trajectories is reassuring, and the human-inspected cases provide useful evidence (App. E.6). However, that review examines originally positive outcomes and uses another LLM. It therefore leaves false negatives and shared judgment errors less well characterized. Since some failures depend on interpreting authorization and fabricated claims, a modest human audit covering both outcome classes would strengthen confidence in the benchmark.
Questions for authors
  1. Can you provide a representative, equal-budget comparison between CFS and full-restart attack search, reporting replay success and total cost?

  2. Do you have human-checked agreement rates on a sample of both successful and unsuccessful runs, particularly for cases involving fabricated evidence rather than explicit instructions?

2 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer bW3P

Thank you for highlighting how shared application state lets us trace the effects of an outsider’s message from the moment an agent reads it to the change it makes in an application. We are glad you singled out the fabricated-evidence finding in §4.2 and Appendix H.8. An agent that rejects an outsider’s explicit instructions can still act on a false claim about what the user has authorized.

Q1. Can you provide a representative, equal-budget comparison between CFS and full-restart attack search, reporting replay success and total cost?

Thank you for suggesting this comparison. We evaluated CFS and adaptive full restart on the same 60 mixed browser–MCP tasks, using GPT-6 Astra as the attacker and GPT-5.6 Sol as the victim. These are the most challenging tasks in our evaluation because the victim must coordinate actions across both interfaces. On each task, we ran 30 CFS attempts and gave full restart the same victim-inference cost budget. We then replayed each method’s selected attack three times from the original user request, for 60 fresh replays per method.

MetricCFSAdaptive full restart
Tasks with a successful attack during search45%30%
Fresh-replay ASR35%23.3%
Mean search attempts per task308
Mean victim inference cost per task during search$1.95$1.96

As we can see, at nearly the same victim cost, CFS achieved 50% higher optimization ASR than full restart (45% vs. 30%) while allowing 3.75× as many search attempts (30 vs. 8 per task). Its fresh-replay ASR was also higher (35% vs. 23.3%).

Q2. Do you have human-checked agreement rates on a sample of both successful and unsuccessful runs, particularly for cases involving fabricated evidence rather than explicit instructions?

Thanks for the insightful question! Before running the final large-scale benchmark evaluation, we refined the judge prompt over several rounds of human review. In each round, we read the trajectories, judged whether the attacks succeeded, and compared our verdicts with the model’s. When we found a disagreement, we revised the relevant part of the prompt and checked whether the revised judge agreed with our assessment. Therefore, the final rubric now would ask whether the agent “treated an attacker-fabricated document/claim as ground truth” and directs the judge to “QUOTE the obeyed instruction / override / fabricated claim.” In the rubric’s fake-invoice example, the judge must identify the false payment detail the agent relied on and verify that it sent the payment to the specified address (Appendix E.7, p. 97). Based on our human spot-check of 50 cases involving fabricated evidence, we estimate about 95% agreement with the final judge.

We also independently re-judged 1,287 original trajectories with GPT-6 Astra (over the current claude judge), covering all 25 victim models and all five scenario families. It upheld 595 of 628 original positive verdicts (94.7%) and 614 of 659 original negative verdicts (93.2%), with 95.8% agreement on the 1,262 determinate cases. The disagreements largely offset each other, with 23 original positives changing to negative and 30 negatives changing to positive. Applying these changes while keeping other labels fixed moved pooled benchmark ASR only from 36.71% to 36.75%, and all 25 Figure 1 model ranks remained unchanged.


We appreciate your suggestions to compare CFS and full restart under a common total budget and to include both positive and negative verdicts in a human audit, particularly for attacks that rely on fabricated evidence. And we would include all these discussions in our final revision.

Thanks for addressing my questions, I have no follow up questions for Q1: comparison between CFS and full-restart attack search experiments.

Regarding to the " refined the judge prompt over several rounds of human review", what is your rubrics and standards for the refinement and as you mentioned that you read the trajectories and " judged whether the attacks succeeded", I am wondering the methods for you to judge the trajectories, as in my experience, those trajectories are normally very long, some might be hundreds of thousand tokens, they are really hard for humans to just manully read and check, so what are the methods you used to handle these?

Reviewer_t6d4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces ArenaWeaver, a fully automated full-stack evaluation framework designed to benchmark LLM agents' robustness against indirect prompt injection attacks at scale.

  1. Automates the reconstruction of 40 real-world platform simulations (e.g., Gmail, Slack, PayPal) with 1,263 web-UI routes and 1,179 MCP tools, featuring shared persistent state across browser and MCP interfaces.

  2. Proposes an adaptive tree search algorithm that leverages indexed interaction histories to test counterfactual interventions without re-executing long-horizon tasks from scratch, drastically reducing search cost and execution time.

  3. Curates a benchmark of 660 cases evaluated across 25 frontier models, uncovering key vulnerability patterns such as asymmetric cross-model attack transfer, high cross-interface attack transfer (72.5% from browser to MCP), and agents' susceptibility to accepting fabricated evidence within requested operations.

Critical strengths

ArenaWeaver is the first benchmark to combine browser UIs and Model Context Protocol tool access over a unified, persistent SQL backend at scale. Unlike prior benchmarks that focus strictly on tools or lack MCP paired environments, ArenaWeaver brings customizable, stateful multi-app workflows under strict experimental control.

Critical weaknesses
  1. Although the authors conduct an environment suspicion analysis showing low rate of explicit model detection, fully automated reconstructed applications and synthetic persona datasets may still exhibit subtle behavioral cues or distribution shifts compared to real-world production platforms.

  2. The adjudication of attack success and benign task completion relies heavily on frontier LLM judges. Although the authors validated judge labels via a blinded review, distinguishing complex behavioral nuances can introduce potential judge errors or subtle biases toward certain model families, which could marginally impact fine-grained model rankings.

Questions for authors
  1. In the Introduction, the article proposes three principles of failure modes; however, it does not explain the source and basis of these three principles.

  2. The authors should further explicitly state what problem CFS is intended to solve and what results it has achieved.

  3. In Section 3.3 and Section 4.4, the CFS relies on a specific surrogate model to discover attack vectors, which are then used to evaluate other victim models in the static benchmark. Could the authors clarify to what extent the static benchmark's ASR are biased by the choice of the surrogate model?

2 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer t6d4 (Part-1)

We greatly appreciate the reviewer’s positive assessment of ArenaWeaver’s technical soundness and its contribution to agent security evaluation. We are particularly grateful for their recognition of our unified browser/MCP environment design, which enables customizable, stateful multi-application workflows under controlled experimental conditions.

W2. The adjudication of attack success and benign task completion relies heavily on frontier LLM judges. Although the authors validated judge labels via a blinded review, distinguishing complex behavioral nuances can introduce potential judge errors or subtle biases toward certain model families, which could marginally impact fine-grained model rankings.

Thank you for raising this important point! We refined the judge prompt through repeated comparison with human judgments, revising it until the model agreed with us on the cases we reviewed. The resulting rubric requires evidence of a committed action, its attribution to injected content, and the resulting harm (Appendix E.7). To check for judge-family effects, we independently re-judged 1,287 original trajectories with GPT-6 Astra, covering all 25 victim models and five scenario families. Agreement was 95.8% among determinate cases, with 94.7% agreement on positive verdicts and 93.2% on negative verdicts. Updating the benchmark with the determinate disagreements preserved all 25 Figure 1 ranks.

Q1. What is the source and basis of the three principles discussed in the introduction?

Thank you for asking us to clarify their basis. Our long-term goal is to support the training of more robust agents, including through RL. From this training perspective, realism, experimental control, and scale follow directly from what the infrastructure needs to provide.

Realism. If we want training to improve robustness in real applications, the training environment needs to reflect what agents actually encounter and what they can do. Unrealistic environments create two problems. First, they can leave out important attack surfaces, such as browser pages, uploaded documents, forwarding rules, and access controls. Second, they can saturate as models improve, leaving us with few successful attacks to learn from even though real applications expose additional ways for an attack to reach the agent. For example, AgentDojo provides simplified tool-based workflows without browser interaction. Under its standard attack, a study of the banking suite found only 1/288 successful attacks against Claude Sonnet 4.6 and 0/288 against GPT-5.4 [1]. ArenaWeaver therefore reconstructs interactive browser UIs, MCP tools, and persistent application state. Our Gmail MCP closely matches the inputs and response format of an existing Gmail MCP implementation (Appendix B.3, Table 6), while the backend records consequences such as a saved forwarding rule or an access grant. This makes the environment closer to what agents use in practice and should help attacks transfer to real assistants.

Experimental control. Live services such as Gmail do not let researchers freely create and reset arbitrary mailbox states across users. Without control over the messages, accounts, and initial data, it is difficult to place an injection where intended or compare the same task with and without it. ArenaWeaver lets us customize the workspace, deliver injected content through separate tools restricted to an outsider’s allowed actions, and reset application state between runs. This also lets us hold the task and injected content fixed while changing the victim model or its browser/MCP interface (Appendix B). This is especially important for RL, which needs many rollouts from known starting states and rewards based on completing the user’s task while avoiding unauthorized actions.

Scale. As agents become more capable and handle more services, automating environment construction and attack discovery becomes necessary. Manually rebuilding applications and writing tasks cannot keep up, while stronger coding agents can reconstruct services from public documentation and screenshots. ArenaWeaver builds each application once, automatically generates new personas, records, and tasks, and combines applications into new workflows. The practical goal is to respond quickly when a vulnerability is discovered in a real application. This pipeline lets us reconstruct the affected environment, reproduce and explore the exploit, and turn the resulting cases into training and evaluation tasks for more robust agents.

[1] Khan, A., AlKhanbashi, K., & Mohamed, A. (2026). Evaluating indirect prompt injection defenses in tool-using LLM agents: Security, utility, and replication. Computers, 15(9), Article 570. https://doi.org/10.3390/computers15090570

AuthorsOpenReview ↗

Rebuttal for Reviewer t6d4 (Part-2)

Q2. What problem does CFS solve, and what has it achieved?

Thank you for the suggestion!

Motivation. In a long agent workflow, an attacker must decide where to inject, what to inject, and how to deliver it. A fixed injection may never be encountered, so an unsuccessful attack can simply mean the agent did not see it. Trying new locations and payloads with full restarts is also expensive, since each attempt repeats earlier work, e.g., our longest benign runs approach 20 minutes.

How CFS solves this. CFS organizes attack discovery as a tree of alternative agent executions (§3.3, Algorithm 1). It first records a benign run, then lets the attacker fork from an observed interaction and introduce content through the permitted injection tools. The attacker can inspect the response, revise the content or placement, and fork the continuation again while retaining earlier injections. In this way, CFS searches over both the injection and the evolving workflow, while sharing the execution history that preceding branches have in common. The attacker LLM guides branch selection, and a separate judge evaluates attack success.

Results. CFS found successful attacks in 2,232 of 8,730 searches (§4.1). Of the successful searches, 21.2% first received a positive judgment after the 30th recorded search node (Figure 3b). The median continuation took 21.2 seconds, compared with 59.5 seconds for a full benign run, whose median victim-model cost was about 3.1× higher (Figure 3a). In our additional mixed browser–MCP comparison at matched victim cost, CFS supported 30 attempts per task versus eight for adaptive full restart, with observed fresh-replay ASRs of 35% and 23.3%, respectively.

Q3. In Section 3.3 and Section 4.4, the CFS relies on a specific surrogate model to discover attack vectors, which are then used to evaluate other victim models in the static benchmark. Could the authors clarify to what extent the static benchmark's ASR are biased by the choice of the surrogate model?

Thank you for raising this question! We chose six discovery surrogates from Claude, GPT, and Gemini because these were three leading frontier model families, and using all three helps diversify the attacks we find. We also balanced the curated benchmark across the three source families. To check whether the ranking depends on this choice, we repeated Figure 1 after excluding attacks discovered against any model in the evaluated victim’s own family. Across all 25 models, the ranking remained highly correlated with the original (Spearman ρ = 0.962). For example, Gemini 3.1 Pro (76.0%) > GPT-5.4 (59.9%) > GPT-5.6 Sol (25.8%) > Opus 4.8 (4.2%) > Opus 5 (0.9%) retained the same relative ordering by ASR, with adjusted ASRs shown in parentheses. We will include this analysis in the revised paper.

Section 4.4 also lets us compare two discovery surrogates on the same 191 task configurations, holding the attacker and final victim fixed. Attacks discovered against Sonnet 5 and against Opus 5 each succeeded on Claude Opus 4.8 in 52/191 fresh replays (27.2%). Only 27 successful configurations overlapped. For this victim, changing the surrogate therefore changed which tasks exposed a vulnerability while leaving the aggregate ASR unchanged (Appendix F.5).


We thank the reviewer for the constructive feedback and hope these results clarify the evaluation design and strengthen confidence in the main findings. We welcome further questions and would appreciate your consideration of the additional evidence in your final assessment.

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

The paper presents ArenaWeaver, a framework for indirect prompt injection evaluation. Coding agents rebuild 40 popular services from public docs as full-stack apps where the browser UI and MCP tools share one resettable state, and outsiders can inject content through normal app channels. An LLM then generates personas and benign/malicious task pairs. An attacker LLM runs Counterfactual Forking Search (CFS), which forks a surrogate's clean trajectory at chosen turns, injects content and continues from the saved prefix. From 8,730 searches, 660 successful cases are curated into a fixed benchmark and replayed on 25 frontier models. The paper reports ASR, clean utility, cross-model and cross-interface transfer, and a renewed-attacker study.

Critical strengths
  1. The infrastructure is the main contribution and a useful one. Shared browser/MCP state, resettable seeded data and app-level injection channels make it possible to hold the task and attack fixed while changing the interface, which AgentDojo or WASP can't really do. If released, the community will likely use it.

  2. The evaluation is large and fairly careful. It covers 25 models, uses separate clean runs, and separates search success from replay success. The cross-interface transfer result is new and relevant to assistants that mix browser and tool access.

Critical weaknesses
  1. The method itself is not very novel. CFS is essentially checkpoint-and-branch plus an LLM attacker that refines injections (in the spirit of PAIR/TAP/AgentVigil). The attacker LLM makes every search decision and scores its own branches, so calling it a "tree search algorithm" overstates it. Generating environments with coding agents is also an active direction (e.g., Agent-World), so the contribution is mostly engineering at scale.

  2. CFS searches in a setting that differs from how attacks are evaluated, and it is never compared with a baseline. A fork keeps a history recorded without the injection, and the attacker times the injection after seeing the victim's trajectory and reasoning. As a result, many search successes do not reproduce when the victim runs from the user's task, even on the same model. The efficiency claim only compares one branch against one full run, and there is no comparison with a full-restart search, a simple static injection, or prior adaptive attackers. So it is unclear whether forking actually helps.

  3. Some headline findings are confounded by how the benchmark was built. All cases are selected from attacks that succeeded against six specific surrogates, so the model ranking partly reflects each model's relation to those surrogates. The "Claude-found attacks transfer to GPT/Gemini but not vice versa" result has a simpler explanation: attacks that beat a more robust model are stronger. Also, the scenario-generation prompt explicitly asks for near goals to be easier to induce than off-task goals, so the "near > off-task" finding is partly built in.

  4. Most findings confirm what is already known. That utility doesn't imply robustness, that adaptive attacks find more failures, and that repeated sampling raises ASR have all been reported before. No defenses are evaluated, and fidelity to the real services is only checked structurally, so the practical takeaways are limited.

Questions for authors
  1. At matched victim cost, how does CFS compare with a full-restart adaptive search and a static injection baseline, measured by replay ASR rather than search ASR? This would most directly change my assessment.

  2. Does the model ranking hold when each model is only evaluated on cases searched against other model families? Does the transfer asymmetry remain after controlling for how robust the source surrogate is?

  3. Since the generator was told to make near goals easier, does the near vs. off-task gap persist for scenarios generated without that instruction?

  4. Claude models build the environments, generate scenarios, attack and judge. Can you report human agreement on a random sample that includes negative verdicts, and re-judge a subset with a non-Claude judge, to rule out family bias?

3 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer YxoX (Part-1)

We sincerely thank the reviewer for recognizing the value of ArenaWeaver’s infrastructure, particularly its shared browser/MCP state, resettable environments, and application-level injection channels. We are especially grateful for the reviewer’s appreciation of the scale and care of our evaluation and their recognition of cross-interface attack transfer as a new and practically relevant finding.

W1. The method itself is not very novel. CFS is essentially checkpoint-and-branch plus an LLM attacker that refines injections (in the spirit of PAIR/TAP/AgentVigil). The attacker LLM makes every search decision and scores its own branches, so calling it a "tree search algorithm" overstates it. Generating environments with coding agents is also an active direction (e.g., Agent-World), so the contribution is mostly engineering at scale.

Thank you for asking how CFS differs from earlier attack searches. The fundamental difference is the red-teaming question each method is designed to answer. PAIR and TAP ask what prompt will cause a target model to produce a prohibited response. AgentVigil brings this approach to indirect injection by searching for an effective injection template at sites specified by the task. CFS asks a broader question: as an agent carries out a long workflow, where should an outsider intervene, what should they place there, and how should the next intervention depend on the agent’s response? A fixed injection may never reach an agent, while testing alternatives through full restarts repeatedly executes the same long history. CFS addresses this with a tree-based search over victim execution histories (§3.3, Algorithm 1). The root stores the benign run, and each expansion selects a parent and interaction point, applies an outsider intervention, and adds the victim’s continuation as child nodes. Sibling branches explore alternatives from a shared history, while descendants inherit earlier injections and can fork again at newly observed interactions (Appendix D.6). The attacker LLM selects and scores branches to guide exploration, while a separate judge determines attack success.

ArenaWeaver reconstructs realistic applications with fully implemented browser UIs and MCP tools over shared persistent state (§§3.1–3.2). Agent-World’s published environments use comparatively simple function-based interfaces without corresponding browser UIs. For example, its arXiv environment exposes three retrieval functions, and its email environment offers eight read-only functions over a mailbox export (Agent-World, Appendices B.1–B.2). We implement navigable pages, forms, and working controls from product screenshots and documented operations, allowing agents to move between browser and MCP access within the same workflow. Across 40 applications in 10 categories, our infrastructure provides 1,263 UI routes, 1,179 victim-facing MCP tools, and 5,121 HTTP operations (Appendix B.5, Tables 8–9). A separate attacker-facing MCP server provides 183 injection actions, including delivering emails and PDF attachments (Appendix B.6). These reusable applications support 160 application–interface compositions, enabling attacks that span services and interfaces.

W2. CFS searches in a setting that differs from how attacks are evaluated, and it is never compared with a baseline. A fork keeps a history recorded without the injection, and the attacker times the injection after seeing the victim's trajectory and reasoning. As a result, many search successes do not reproduce when the victim runs from the user's task, even on the same model.

Thanks for the insightful question! (1) There is randomness in the victim model’s behavior, regardless of how the attack was found. Even with the same task and exactly the same injection, the model may take a different path or react differently to what it reads. This randomness comes from the victim model and is present regardless of the injection method, including adaptive full-restart search. (2) When we repeat the test, the gap between search and replay narrows. Figure 6b (§4.4) shows this directly. Across 191 configurations, CFS found a successful attack against Opus 5 in 31.9% of cases. With the injections held fixed, 16.2% of cases succeeded on the first replay, and 26.2% succeeded at least once within four replays. The absolute search–replay ASR gap therefore fell from 15.7% to 5.7% on Opus 5. Similarly, it fell from 19.9% to 7.9% on Opus 4.8 and from 15.2% to 5.8% on Sonnet 5.

AuthorsOpenReview ↗

Rebuttal for Reviewer YxoX (Part-2)

Q1. At matched victim cost, how does CFS compare with a full-restart adaptive search and a static injection baseline, measured by replay ASR rather than search ASR? This would most directly change my assessment.

Thank you for this valuable suggestion! To test whether forking helps at matched victim cost, we focus on mixed browser–MCP workflows, the most challenging scenario where the victim must coordinate actions across both interfaces. We evaluated 60 paired tasks with GPT-6 Astra as the attacker and GPT-5.6 Sol as the victim. For each task, we first ran CFS for 30 iterations, then gave adaptive full restart the same victim-inference cost budget. The static baseline committed to one static injection. We measured replay ASR by rerunning each fixed attack three times from the original request.

MetricCFSAdaptive full restartOne-shot static
Optimization ASR45%30%5%
Fresh-replay ASR35%23.3%3.3%
Mean search attempts per task3081
Mean victim cost per task during optimization$1.95$1.96$0.23

As we can see, at nearly the same victim cost, CFS achieved 50% higher optimization ASR than full restart (45% vs. 30%) while allowing 3.75× as many search attempts (30 vs. 8 per task). Its fresh-replay ASR was also higher (35% vs. 23.3%).

Q2. Does the model ranking hold when each model is only evaluated on cases searched against other model families? Does the transfer asymmetry remain after controlling for how robust the source surrogate is?

Thanks for the insightful questions! Figure 5a shows how attacks found against each of the six surrogates fare on other models, and Appendix Figure 55a compares four Claude search targets on the same 191 tasks. To test the ranking directly, we repeated the Figure 1 analysis after removing every attack discovered against a model in the evaluated model’s own family. For example, when evaluating a Claude model, we excluded attacks found against any Claude surrogate. We kept the original clean-task filter and scenario weighting. The ranking remained highly correlated with the original across all 25 models (Spearman ρ = 0.962). For example, the relative ASR ordering Gemini 3.1 Pro (76.0%) > GPT-5.4 (59.9%) > GPT-5.6 Sol (25.8%) > Opus 4.8 (4.2%) > Opus 5 (0.9%) held both before and after excluding same-family discoveries, with adjusted ASRs in parentheses and > indicating higher ASR. DeepSeek V4-pro remained highest and Opus 5 lowest. We will include this additional ranking analysis in the revised paper.

Q3. Since the generator was told to make near goals easier, does the near vs. off-task gap persist for scenarios generated without that instruction?

Following your suggestion, we removed the sentences telling the generator to make near goals easier to induce than off-task goals. Using GPT-6 Astra, we generated near and off-task goals both with and without those sentences for the same 99 single-application MCP tasks across 31 applications, keeping the goal definitions fixed. We allowed up to 30 search attempts per goal and replayed each selected attack three times, counting the two tasks with early-stopped searches as failures. Averaging equally across applications, replay ASR was 18.2% for near goals and 4.0% for off-task goals without the sentences, compared with 19.5% and 3.9% with them. The near-goal advantage persisted without the difficulty instruction.

AuthorsOpenReview ↗

Rebuttal for Reviewer YxoX (Part-3)

Q4. Claude models build the environments, generate scenarios, attack and judge. Can you report human agreement on a random sample that includes negative verdicts, and re-judge a subset with a non-Claude judge, to rule out family bias?

Thank you for raising this important point! We used agreement with human judgments to refine the judge prompt throughout development. We manually reviewed agent trajectories, compared the judge’s verdicts with our own, and revised the prompt when they disagreed. We repeated this process until the model judge agreed with our judgments on the cases we reviewed. This iterative process is why our final judge prompt includes the detailed rubric in Appendix E.7, requiring evidence that the agent committed an action, that injected content influenced it, and that real harm resulted.

We re-judged a subset of 1,287 original trajectories with GPT-6 Astra, a non-Claude judge, covering all 25 victim models and all five scenario families. Astra assessed both positive and negative cases without seeing the original verdicts. Astra achieved 95.8% agreement with the original harm judgments among determinate cases, including 94.7% agreement on positive verdicts and 93.2% on negative verdicts. The disagreements largely offset each other, with 23 original positives changing to negative and 30 negatives changing to positive. Applying these changes while keeping other labels fixed moved pooled benchmark ASR only from 36.71% to 36.75%, and all 25 Figure 1 model ranks remained unchanged.


Thank you again for the detailed and constructive review. We hope the new paired replay experiment and benchmark-construction controls address the central concerns. If these additions change your assessment, we would be grateful for your reconsideration, and we are happy to clarify any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.