acceptodds
Under review as a conference paper at ICLR 2027

AutoRedClaw: Automated Red-Teaming for Claw-Style Agents

Abstract

Agent systems are shifting from task-scoped tool-use agents to persistent, personalized assistants such as OpenClaw. These claw-style agents maintain configuration files (e.g., AGENTS.md), long-term memory, and reusable skills. Once poisoned, this persistent state is repeatedly reloaded into the agent's context, creating persistent threats across tasks and sessions. Existing security evaluation systems struggle to evaluate these complex claw-style systems since they rely heavily on manually constructed test cases and environments, which limits their scale and coverage; In addition, these manually designed cases quickly become outdated as agents evolve. To address this, we introduce AutoRedClaw, an automated red-teaming framework that generates end-to-end attacks together with the environments needed to execute them. AutoRedClaw combines a strict quality-checking protocol with execution-based validation to ensure the correctness, realism, and effectiveness of generated samples, and adopts task-driven environment construction to keep costs low. To achieve broad coverage, it organizes generation along risk categories, attack strategies, attack stages, and attack surfaces, and uses experience-guided exploration to steer generation toward under-explored regions. AutoRedClaw achieves an effective attack rate above 30% at less than $0.10 per successful sample. From its outputs, we construct RedClaw-Bench, a 300-case benchmark that yields substantially higher attack success rates (ASRs) than AgentDojo, AgentDyn, and DTap. Across seven harnesses and four frontier models, most agent systems exceed 50% ASR, showing that our red-teaming pipeline provides a strong evaluation signal even for the most advanced agent systems and demonstrating the effectiveness of our approach. Our code and benchmark are released at https://anonymous.4open.science/r/AutoRedClaw-DD8E.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
indirect prompt injection, red-teaming, agent security

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_JEKGOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a framework of automated red-teaming for claw-style agents. Specifically, the authors have pointed out that existing security evaluation systems struggle to evaluate these complex claw-style systems since they rely heavily on manually constructed test cases and environments. To address the above-mentioned point, the authors have proposed AutoRedClaw as an automated red-teaming framework that generates end-to-end attacks alongside the environments on which they are executed. Moreover, the authors have constructed RedClaw-Bench as a 300-case benchmark. The experimental results show that the constructed benchmark yields substantially higher attack success rates than existing benchmarks.

Critical strengths
  1. The authors have identified three limitations of prior security benchmarks, namely, high construction costs severely restrict scale and coverage, static benchmarks rapidly become outdated and incomplete coverage of the attack chain.
  2. The authors have proposed AutoRedClaw as an automated red-teaming framework for constructing and executing security evaluations of diverse claw-style agents.
  3. The authors have constructed RedClaw-Bench as a 300-case benchmark that yields substantially higher attack success rates than existing benchmarks, namely, AgentDojo, AgentDyn and DTap.
Critical weaknesses
  1. The contributions of this paper are not clearly summarised in Section 1.
  2. This paper lacks an independent section on related works so it is unclear how AutoRedClaw differs from other existing frameworks.
  3. The experimental evaluation seems insufficient. Specifically, Section 4 lacks comparisons between AutoRedClaw and other baseline frameworks. Moreover, the ablation studies against the effectiveness of the components of AutoRedClaw are missing. In addition, evaluation of the effectiveness of RedClaw-Bench is mainly based on attack success rate, which does not seem to be sufficient, e.g., it is also worth to consider how RedClaw-Bench impacts the computational costs.
Questions for authors
  1. Please highlight the contributions of this paper in an explicit manner.
  2. Please justify how how AutoRedClaw differs from other existing frameworks.
  3. Please address the points on the experimental evaluation, i.e., baseline comparison, ablation studies and effectiveness of RedClaw-Bench.
7 comments
AuthorsOpenReview ↗

Response to Reviewer JEKG [1/4]

We thank the reviewer JEKG for the patient and professional review, as well as the valuable suggestions for improvement. Our responses to each questions are below. We have added these new experiments and discussions to our revised version (marked in blue).

W1&Q1 : The contributions of this paper are not clearly summarised in Section 1.

A1: We are sorry for the confusion. To clarify the contributions of this paper, we summarize two core limitations of existing work, and how AutoRedClaw solves each of them.

Limitation 1: Limited scalability of manual construction. Existing benchmarks heavily rely on human experts to build every task and environment by hand. So the cost is high, the scale is small, and the benchmark is easy to be outdated. This problem is getting more serious now, since agent systems update really fast and new harnesses and models keep coming out, and a manually built benchmark can not catch up with them.

AutoRedClaw Solution: We solve this from two sides. First, AutoRedClaw makes the whole red-teaming process automatic. It generates the tasks, the attacks and the executable environments without manual effort (Sec. 3.1.2, 3.2), so the scale is no longer limited by human experts. When a new model or harness comes out, our adapter design can be easily extended to these new agents, and we can simply generate the cases again. Second, we significantly reduce the cost of the automatic generation. Our three-stage generation design first generates a high-level plan of each case and then fills in the details in the materialization stage, which makes sure that even with a cheap model as the generator, it can still efficiently find successful attack cases for advanced models at a very low cost. The probe stage further reduces this cost by more than 50%. The automated pipeline and the low cost make AutoRedClaw an effective and practical solution for evaluating current and future advanced agent systems.

Limitation 2: Limited exploration ability of the existing red-teaming paradigm. Some works already use automated red-teaming to reduce the manual effort, but they still have a limitation in what they can explore. The goal of red-teaming is to discover the risks that we do not know yet. However, most existing methods start from a task and an environment that are already defined, and they only optimize the payload inside this fixed setting. So they can only test the predefined risk scenarios, and what they find is just a stronger attack on a known scenario. It is still not a new risk of the agent. This limitation is more serious for advanced agent systems, since their risks largely depend on what task the agent is doing, what it can touch in the environment and what is kept for later tasks. For example, a persistent-state attack needs a state file to be poisoned and a later task to trigger it, and this can never be found by optimizing a payload inside one fixed task.

AutoRedClaw Solution: AutoRedClaw proposes a new red-teaming paradigm to solve this. It explores the tasks, environments and attacks together, so the scenario itself also becomes a part of the search. To make sure the exploration covers enough risks, we combine two kinds of guidance. First, the rich dimensions give the exploration a wide starting space. We define 2 attack stages, 5 surfaces, 9 risk categories and 12 attack strategies from the well-defined taxonomies in agent security, so the generation can cover the known risk types from the beginning. Second, the experience-guided exploration keeps expanding this space during generation. It lets the generator first go to the directions that are not explored yet, so new scenarios keep coming out and the generation does not collapse to several similar ones. With the rich dimensions as the basis and the experience as the driver, AutoRedClaw can keep uncovering new threats to advanced agent systems.

We have reorganize the introduction following these two points, and list our contributions in the same way.

AuthorsOpenReview ↗

Response to Reviewer JEKG [2/4]

W2&Q2 : This paper lacks an independent section on related works, so it is unclear how AutoRedClaw differs from other existing frameworks.

A2: We are sorry for the confusion. We put the Related Work section in Appendix B due to the page limit, and we will move it to the main text in the revised version. To the best of our knowledge, there is no existing large-scale red-teaming work that jointly explores the tasks, environments and attacks for agent systems. Some works also generate the task together with the environment, but they are not in the security scope. Most previous red-teaming works start from a task and an environment that are already defined, and they repeatedly optimize a payload inside this fixed setting, so they do not generate the whole attack together with the environment from scratch. This is exactly the difference between AutoRedClaw and existing frameworks.

To better show the difference and the advantages of AutoRedClaw over existing related work, we conduct experiments against existing environment generation and red-teaming works. Since no single method jointly explores the tasks, environments and attacks for agent systems as AutoRedClaw does, we compare with the two related lines separately.

For environment generation, we employ AutoControl Arena (ACA) [3] as the baseline, which also uses LLMs to synthesize executable test environments. We compare the diversity of its benchmark X-Bench (70 scenarios) with that of the cases generated by AutoRedClaw. We employ the Vendi Score [1] as the diversity metric, which roughly measures how many truly different samples a set contains. Since the Vendi Score grows with the number of samples, we also generate 70 cases with AutoRedClaw. The embedding model is all-mpnet-base-v2.

For red-teaming, we employ AutoDojo [2] as the baseline, which is an iterative red-teaming framework built on top of AgentDojo, and we compare the average cost of finding one successful attack sample. We use GPT-5.4 as the victim model, and AutoRedClaw runs on Codex. For all other settings, we follow the official configuration of AutoDojo. The cost includes all stages for both methods.

The results are shown below:

Environment GenerationVendi Score ↑Red-teamingAverage Cost per Success ($) ↓
AutoControl Arena29.0AutoDojo1.037
AutoRedClaw31.4AutoRedClaw0.176

We observe that on environment generation, AutoRedClaw achieves a Vendi Score of 31.4, which is comparable to and slightly higher than that of ACA (29.0). This shows that our automatically generated tasks are competitive with an advanced task-environment generation framework, even though all of our samples focus on claw-style agents. On red-teaming, AutoRedClaw finds one successful attack sample at about 5.9 times lower cost than AutoDojo (\0.176vs.0.176 vs. \\1.037). This holds even though AutoDojo uses a much simpler harness and borrows its environments from AgentDojo, which are also generally simpler than ours.

To sum up, AutoRedClaw fills the gap of jointly exploring the tasks, environments and attacks for advanced agent systems, and it shows strong quality and effectiveness in both environment generation and attack generation.

AuthorsOpenReview ↗

Response to Reviewer JEKG [3/4]

W3&Q3 : The experimental evaluation seems insufficient, including baseline comparison, ablation studies, and the effectiveness of RedClaw-Bench beyond ASR (e.g., computational costs).

A3: Thank you for your constructive suggestions. We conduct additional experiments for all these points.

Baseline comparison. We have conducted the baseline comparison experiments in A2. Since no existing method jointly explores the tasks, environments and attacks for agent systems, we compare AutoRedClaw with the two related lines separately. For environment generation, we employ AutoControl Arena as the baseline and compare the diversity of the generated scenarios with the Vendi Score. For red-teaming, we employ AutoDojo as the baseline and compare the average cost of finding one successful attack sample, with GPT-5.4 as the victim model. As we can see, AutoRedClaw gets a higher Vendi Score than AutoControl Arena (31.4 vs. 29.0), and finds one successful attack at about 5.9 times lower cost than AutoDojo (\0.176vs.0.176 vs. \\1.037). It shows that AutoRedClaw is effective on both environment generation and attack generation.

Ablation study. We ablate two components, the experience-guided exploration and the probe. For the experience-guided exploration, we generate 100 samples with the diversity guidance and 100 samples without it, and all other settings are kept the same. For the probe, we compare the generation with and without it, with deepseek-v4-pro as the probe model and Codex + GPT-5.4 as the target agent. The results are shown below:

SettingVendi Score ↑ASRCost per Success (\$) ↓
w/o diversity guidance13.133.0%-
w/ diversity guidance32.838.0%-
w/o probe-37.6%0.406
w/ probe-32.8%0.176

We observe that removing the diversity guidance makes the Vendi Score drop from 32.8 to 13.1 (p < 0.001), so the generation collapses to much fewer scenarios without it. As a reference, 100 totally unrelated texts only get 44.1, so our samples reach 74% of this upper bound, while those without the guidance only reach 30%. For the probe, the ASR only drops slightly from 37.6% to 32.8%, but the cost per successful attack drops from \0.406to0.406 to \\0.176, which is about 2.3 times lower. This is because the probe filters out the broken cases before they reach the costly target agent, and it does not throw away many cases that can really attack the target, even though the probe and the target come from different model families.

Effectiveness beyond ASR. Since AutoRedClaw mainly focuses on the persistent state, we add the standard indirect prompt injection (IPI) surface to AutoRedClaw generation pipeline, which has the same threat model as existing benchmarks like AgentDojo and DTap. On this surface, AutoRedClaw gets an ASR of 38.8% on OpenClaw + DeepSeek-V4-Pro, and the cost of finding one successful attack is only about \$0.036. To evaluate the case quality, we take 100 cases generated on this surface and all the 107 IPI cases from the workflow domain of DTap. We employ three LLM judges from different vendors (GPT-6.1-Sol, Gemini-3.1-Pro and Claude-Opus-5.5) and three more human evaluators. Each judge sees a neutral case card without its source, and scores every case once from 1 to 5 on 8 dimensions and gives an overall realism score. The averaged results are shown below:

DimensionAutoRedClaw (LLM)DTap (LLM)AutoRedClaw (human)DTap (human)
Benign task realism4.803.354.693.13
Environment fidelity4.052.043.761.89
Injection carrier naturalness4.793.264.683.36
Injection voice and stealth4.312.643.852.40
Attacker plausibility3.642.403.362.42
Harm severity3.773.643.733.47
Verification soundness4.423.164.112.64
Goal conflict4.413.594.483.50
Overall realism3.852.173.552.24

We observe that AutoRedClaw is significantly better on 7 of the 8 dimensions (p < 0.001), and the overall realism is 3.85 vs. 2.17 under the LLM judges and 3.55 vs. 2.24 under the human evaluators. The verification soundness, which shows whether the success criterion really measures the harm, is 4.42 vs. 3.16, and the judges think a real attacker would do this in 82% of our cases, compared to 38% for DTap. Harm severity is similar on the two benchmarks, so the higher score is not because our attacks are more harmful.

Reviewer_L2FLOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AutoRedClaw is an automated security evaluation framework for persistent agents such as OpenClaw. It generates benign tasks, malicious instructions, and executable environments to assess whether untrusted content can poison an agent’s persistent state and whether that state subsequently induces harmful behavior. The authors also curate a 300-case benchmark, RedClaw-Bench, from the generated cases.

Critical strengths

A strength of this work is that it automates the creation of attacks targeting the persistent state of claw-style agents, making it easier to uncover vulnerabilities in their configuration files, memory, and skills. The authors also release a 300-case benchmark, RedClaw-Bench, and provide adapters that allow the same cases to be tested across different agent frameworks.

Critical weaknesses
  1. I think the title suggests a somewhat broader security evaluation than the paper actually provides. The experiments mainly focus on poisoning persistent state and triggering harmful behavior from it, so I would make that scope clearer.

  2. The introduction lays out the limitations of prior work, but the discussion feels a little long, while the contributions are harder to follow. It would help to say more directly which part of the framework addresses each limitation.

  3. The paper argues that agent systems keep changing, but generation is still guided by predefined categories and attack surfaces. Experience-guided exploration improves coverage within that space; I am less clear on how the framework would identify a new type of component outside it.

  4. I appreciate the effort behind the automated pipeline, but I would like a sharper account of its methodological novelty relative to existing work on attack and environment generation. The paper compares benchmark results with AgentDojo, AgentDyn, and DTap, but a comparison with a closely related generation method under similar conditions would be more informative.

  5. The reported cost is attractive, and task-driven environment construction offers a plausible explanation. Still, without a cost breakdown or a controlled comparison, it is hard to tell how much of the saving comes from that design and how much comes from the choice of DeepSeek models.

  6. Inject and Trigger are evaluated separately: Trigger begins with a payload already in place. I would like to see the success rate for the full sequence on the same attack, from external injection to harm in a later task. That would provide more direct support for the paper’s end-to-end claim.

Questions for authors
  1. The harness adapters appear to map a shared set of persistent surfaces and capabilities across agents. Does this mean that the evaluation primarily covers features shared with OpenClaw, while agent-specific components may be missed? For example, if another agent has a distinct memory mechanism or persistent component, how would AutoRedClaw identify and generate attacks for it, rather than simply transfer the existing cases?

  2. Across the agents you tested, did you observe any attack pattern that was consistently successful regardless of the harness or underlying model? I would be interested in what that pattern suggests about vulnerabilities shared by these systems.

  3. The paper finds that SOUL.md is relatively easy to poison during Inject, whereas AGENTS.md is especially effective during Trigger. What do you think drives this difference—is it mainly how agents protect these files, or how much authority they assign to their contents?

10 comments
AuthorsOpenReview ↗

Response to Reviewer L2FL [1/7]

We thank the reviewer L2FL for the patient and professional review, as well as the constructive suggestions for improvement. Our detailed responses for each questions are below. We have added these new experiments and discussions to our revised version (marked in blue).

W1 : The title suggests a somewhat broader security evaluation than the paper actually provides.

A1: Thank you for the valuable suggestion. We agree that our evaluation mainly focuses on the persistent state of claw-style agents, and we address this in two ways. First, we will revise the title and the abstract to make the scope clear, e.g., "AutoRedClaw: Automated Red-Teaming for Persistent-State Attacks on Claw-Style Agents".

Second, we add the environment surface to AutoRedClaw, which is the standard indirect prompt injection (IPI) setting used in traditional benchmarks such as AgentDojo and DTap. On this surface, the attacker can only control the external content that the task needs, such as web pages, emails and files, and the harm happens in the same task, so the threat model is the same as that of existing benchmarks. On this surface, AutoRedClaw achieves an ASR of 38.8% with OpenClaw + DeepSeek-V4-Pro as the target agent, which means that for every 100 samples it generates, AutoRedClaw finds about 39 that successfully attack this agent. On average, the total cost of finding one successful attack sample is only about \$0.036.

To evaluate the quality of the cases on this surface, we generate 100 cases and compare them with all the 107 IPI cases in the workflow domain of DTap. Each case is converted into the same neutral case card, which contains the user task, what the agent can see in the environment, the injected content and where it appears, the attacker goal and the success criterion, without any information about its source. We employ three LLM judges from different vendors, i.e., Group A (GPT-6.1-Sol, Gemini-3.1-Pro and Claude-Opus-5.5), and three human evaluators in Group B. The evaluation guidelines are given in a separate comment ([7/7]) due to the length limit.

The overall realism scores are shown below:

GPT-6.1-SolGemini-3.1-ProClaude-Opus-5.5Group A Avg.Human 1Human 2Human 3Group B Avg.
DTap2.412.211.902.172.202.172.362.24
AutoRedClaw3.684.353.513.853.463.573.623.55

We observe that all six judges consistently give our cases a higher overall realism score than DTap (p < 0.001 for every judge). Therefore, AutoRedClaw also covers the threat model of existing benchmarks with higher case quality, and the persistent surfaces are the new part we add on top of it.

AuthorsOpenReview ↗

Response to Reviewer L2FL [2/7]

W2 : The introduction lays out the limitations of prior work, but the discussion feels a little long, while the contributions are harder to follow.

A2: Thank you for the constructive suggestion. We agree that the discussion of prior work is a little long, and we will shorten it. In fact, all the limitations we discussed come from two core problems of existing work, and AutoRedClaw is designed to solve both of them.

Limitation 1: Limited scalability of manual construction. Existing benchmarks heavily rely on human experts to build every task and environment by hand. So the cost is high, the scale is small, and the benchmark is easy to be outdated. This problem is getting more serious now, since agent systems update really fast and new harnesses and models keep coming out, and a manually built benchmark can not catch up with them.

AutoRedClaw Solution: We solve this from two sides. First, AutoRedClaw makes the whole red-teaming process automatic. It generates the tasks, the attacks and the executable environments without manual effort (Sec. 3.1.2, 3.2), so the scale are no longer limited by human experts. When a new model or harness comes out, our adapter design can be easily extended to these new agents, and we can simply generate the cases again. Second, we significantly reduce the cost of the automatic generation. Our three-stage generation design first generates a high-level plan of each case and then fills in the details in the materialization stage, which makes sure that even with a cheap model as the generator, it can still efficiently find successful attack cases for advanced models at a very low cost. The probe stage further reduces this cost by more than 50%. The automated pipeline and the low cost make AutoRedClaw an effective and practical solution for evaluating current and future advanced agent systems.

Limitation 2: Limited exploration ability of the existing red-teaming paradigm. Some works already use automated red-teaming to reduce the manual effort, but they still have a limitation in what they can explore. The goal of red-teaming is to discover the risks that we do not know yet. However, most existing methods start from a task and an environment that are already defined, and they only optimize the payload inside this fixed setting. So they can only test the predefined risk scenarios, and what they find is just a stronger attack on a known scenario. It is still not a new risk of the agent. This limitation is more serious for advanced agent systems, since their risks largely depend on what task the agent is doing, what it can touch in the environment and what is kept for later tasks. For example, a persistent-state attack needs a state file to be poisoned and a later task to trigger it, and this can never be found by optimizing a payload inside one fixed task.

AutoRedClaw Solution: AutoRedClaw proposes a new red-teaming paradigm to solve this. It explores the tasks, environments and attacks together, so the scenario itself also becomes a part of the search. To make sure the exploration covers enough risks, we combine two kinds of guidance. First, the rich dimensions give the exploration a wide starting space. We define 2 attack stages, 5 surfaces, 9 risk categories and 12 attack strategies from the well-defined taxonomies in agent security, so the generation can cover the known risk types from the beginning. Second, the experience-guided exploration keeps expanding this space during generation. It lets the generator first go to the directions that are not explored yet, so new scenarios keep coming out and the generation does not collapse to several similar ones. With the rich dimensions as the basis and the experience as the driver, AutoRedClaw can keep uncovering new threats to advanced agent systems.

We have reorganize the introduction following these two points, and list our contributions in the same way.

AuthorsOpenReview ↗

Response to Reviewer L2FL [3/7]

W3 : Generation is still guided by predefined categories and attack surfaces, and experience only improves coverage within that space. How would AutoRedClaw discover new components outside this space?

A3: We would like to clarify that the predefined risk categories, attack strategies and surfaces only guide the generation on a few dimensions, because these are well-defined taxonomies in current agent security. It does not mean the generator is limited by predefined categories on the other dimensions. For example, we do not give any predefined category for the scenarios, and with the help of experience, the generator easily expands to dozens of different scenarios during long-term generation, from devops and finance to customer support and HR.

To show how much the experience helps, we generate 100 samples with the diversity guidance and 100 samples without it, and all other settings are kept the same. We employ the Vendi Score [1] as the diversity metric, which roughly means how many different clusters (scenarios) a set really has. The embedding model is all-mpnet-base-v2. The results are shown below:

SettingVendi Score ↑ASR
w/o diversity guidance13.133.0%
w/ diversity guidance32.838.0%

We observe that the diversity guidance increases the Vendi Score from 13.1 to 32.8 (p < 0.001). As a reference, 100 totally unrelated texts only get a Vendi Score of 44.1 (equal to upper bound) under the same embedding model, so the samples with the diversity guidance reach 74% of this upper bound, while those without it only reach 30%. It shows that the generator is not stuck in a fixed space and really explores many new scenarios with the help of experience.

Moreover, if a new category appears on these predefined dimensions in the future, eg., a new attack strategy or a new risk category, we only need to add a short description of it into the taxonomy, and the whole pipeline can be directly reused without any other change.


W4 : A sharper account of the methodological novelty relative to existing work on attack and environment generation. A comparison with a closely related generation method under similar conditions would be more informative.

A4: Thank you for the valuable feedback. To the best of our knowledge, there is no existing large-scale agent red-teaming work that jointly explores the tasks, environments and attacks for agent systems, so we cannot find a single baseline to compare with as a whole. Some works also generate the task together with the environment, but they are not in the security scope. Most previous red-teaming works start from a task and an environment that are already defined, and they repeatedly optimize a payload inside this fixed setting, so they do not generate the whole attack together with the environment from scratch. This is exactly the problem we aim to solve in this work.

Since no single method shares our setting, we compare with the two related lines separately.

For environment generation, we employ AutoControl Arena (ACA) [3] as the baseline, which also uses LLMs to synthesize executable test environments. We compare the diversity of its benchmark X-Bench (70 scenarios) with that of the cases generated by AutoRedClaw. We employ the Vendi Score [1] as the diversity metric, which roughly measures how many truly different samples a set contains. Since the Vendi Score grows with the number of samples, we also generate 70 cases with AutoRedClaw. The embedding model is all-mpnet-base-v2.

For red-teaming, we employ AutoDojo [2] as the baseline, which is an iterative red-teaming framework built on top of AgentDojo, and we compare the average cost of finding one successful attack sample. We use GPT-5.4 as the victim model, and AutoRedClaw runs on Codex. For all other settings, we follow the official configuration of AutoDojo. The cost includes all stages for both methods.

The results are shown below:

Environment GenerationVendi Score ↑Red-teamingAverage Cost per Success (\$) ↓
AutoControl Arena29.0AutoDojo1.037
AutoRedClaw31.4AutoRedClaw0.176

We observe that on environment generation, AutoRedClaw achieves a Vendi Score of 31.4, which is comparable to and slightly higher than that of ACA (29.0). This shows that our automatically generated tasks are competitive with an advanced task-environment generation framework, even though all of our samples focus on claw-style agents. On red-teaming, AutoRedClaw finds one successful attack sample at about 5.9 times lower cost than AutoDojo (\0.176vs.0.176 vs. \\1.037). This holds even though AutoDojo uses a much simpler harness and borrows its environments from AgentDojo, which are also generally simpler than ours.

To sum up, AutoRedClaw fills the gap of jointly exploring the tasks, environments and attacks for advanced agent systems, and it shows strong quality and effectiveness in both environment generation and attack generation.

Reviewer_FYEBOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This submission proposes the AutoRedClaw. It is an automated red-teaming framework for openclaw-style AI agents with long-term memory, configuration files, and some reusable skills. Compared to traditional agent safety benchmarking, AutoRedClaw not only automatically generates attack tasks but also generates corresponding executable environments and evaluation criteria. It can reduce the cost of manual effort for building a safety benchmark. The experiment shows that many existing agent systems can still be successfully attacked by persistent-state security risks.

Critical strengths
  1. This paper aims to solve an important safety challenge (persistent agent states). It evaluates both the establishment of persistent compromise and its downstream outcomes by separating injection and triggering.

  2. It provides an end-to-end pipeline to automatically generate some attack tasks, executable environments, and evaluation criteria. This framework is practically important for safety evaluations beyond manually generated benchmarks.

  3. The paper tests 7 agent harnesses, not just a single harness, which enhances the generalization of the attack framework. It also provides useful insights into the security and utility trade-offs of different harnesses.

Critical weaknesses
  1. This submission needs an ablation study to demonstrate the improvement contribution of each component.

  2. Selection bias introduced by the validation probe. This paper keeps a candidate as an evaluation case only when the probe agent successfully executes the attack. Therefore, it may favor attack patterns effective against the probe model.

  3. The comparison with other benchmarks is not fully controlled. They differ in threat models, attack surfaces, and success criteria. So, higher ASR on RedClaw-Bench does not necessarily establish better benchmark quality.

Questions for authors

Please refer to the weakness section.

6 comments
AuthorsOpenReview ↗

Response to Reviewer FYEB [1/4]

We thank the reviewer FYEB for the patient and professional review, as well as the valuable suggestions for improvement. Our responses to the reviewer's questions are below. We have added these new experiments and discussions to our revised version (marked in blue).

W1 : This submission needs an ablation study to demonstrate the improvement contribution of each component.

A1: Thanks for the valuable suggestion. In AutoRedClaw, there are multiple components, such as generator, materializer, experience-guided mechanism, probe, state-based judgement, etc. But not each component is suitable to be ablated. The generator, materializer, and the state-based judgment are all the necessary components of AutoRedClaw, AutoRedClaw cannot perform without these basic components. We therefore select experience-guided mechanism, probe as the additional ablation target.

Experience-guided Ablation

To explore the effectiveness of experience-guided mechanism, we compare the generation with it and without it, and employ the Vendi Score [1] as the diversity metric, which roughly means how many different clusters (scenarios) a set really has. We separately generate 100 samples for two settings. The embedding model is all-mpnet-base-v2. The results are shown below:

SettingVendi Score ↑ASR
w/o diversity guidance13.133.0%
w/ diversity guidance32.838.0%

We observe that the diversity guidance brings a significant improvement in diversity, increasing the Vendi Score from 13.1 to 32.8 (p < 0.001), which indicates the effectiveness of the coverage-guided scenario construction and experience-guided exploration in our approach. Moreover, 100 totally unrelated texts only achieve a Vendi Score of 44.1 under the same embedding model. Therefore, the samples generated with the diversity guidance already reach 74% of this upper bound, while those without it only reach 30%. We also observe the same trend under other embedding models and text fields, further demonstrating the robustness of this conclusion.

Probe Ablation

To further explore the effectiveness of the probe, which is designed to save cost and improve efficiency, we compare the generation with the probe and without the probe. We utilize deepseek-v4-pro as the probe model, and Codex + GPT-5.4 as the victim agent. For the with-probe and without-probe settings, we compute the per-sample cost of the probe stage and the evaluation stage, the total per-sample cost, the ASR (using all generated samples as the denominator), and the efficiency (cost per successful attack). The results are shown below:

ModeProbe $/sampleEval $/sampleTotal $/sampleASREfficiency ($/success)
w/o probe—0.1470.14737.6%0.391
w/ probe0.0080.0440.05232.8%0.159

We can observe that even when we use models from different model families as the probe (deepseek) and the victim (GPT), changing from the without-probe setting to the with-probe setting only slightly decreases the ASR from 37.6% to 32.8%, but significantly lowers the cost per successful attack, reducing the cost by approximately 60% (\0.391−>0.391->\\0.159) on this expensive victim model. This demonstrates that the probe does not sacrifice the potential successful attack samples, and also does not introduce too much attack pattern bias toward the probe model, but effectively reduces the red-teaming cost, so it is a practical and effective component in a real-world red-teaming framework.

AuthorsOpenReview ↗

Response to Reviewer FYEB [2/4]

W2 : Selection bias introduced by the validation probe. It may favor attack patterns effective against the probe model.

A2: Thank you for your insightful feedback. Actually, the probe in AutoRedClaw is primarily designed to save the evaluation cost by filtering out the samples with structural issues, eg., the oracle is incomplete, or the attack is not reachable in the benign task. But as you mentioned, this filtering may introduce a bias, since the selected samples may follow the attack patterns that work well on the probe model.

To figure out how much bias the probe introduces, and whether it impacts the attack on the target agent, we conduct a comparison between with probe and without probe (same as A1), with deepseek-v4-pro as the probe model and Codex + GPT-5.4 as the target agent. The results are shown below:

ModeProbe $/sampleEval $/sampleTotal $/sampleASREfficiency ($/success)
w/o probe—0.1470.14737.6%0.391
w/ probe0.0080.0440.05232.8%0.159

We observe that even when we use models from different model families as the probe (deepseek) and the victim (GPT), changing from the without-probe setting to the with-probe setting only slightly decreases the ASR from 37.6% to 32.8%, but significantly lowers the cost per successful attack, reducing the cost by approximately 60% (\0.391−>0.391->\\0.159) on this expensive victim model. This demonstrates that the probe does not introduce too much attack pattern bias toward the probe model, but effectively reduces the red-teaming cost, so it is a practical and effective component in a real-world red-teaming framework.

AuthorsOpenReview ↗

Response to Reviewer FYEB [3/4]

W3 : The comparison with other benchmarks is not fully controlled. They differ in threat models, attack surfaces, and success criteria.

A3: Thank you for the valuable feedback. We agree that ASR alone is a limited criterion for evaluating benchmark quality, since our work and previous benchmarks (e.g., AgentDojo and DTap) differ a lot in their threat models. To make the comparison fairer, we add a new generation surface to AutoRedClaw, the standard indirect prompt injection (IPI) attack, which has the same threat model as AgentDojo and DTap. On this surface, AutoRedClaw achieves an ASR of 38.8% with OpenClaw + DeepSeek-V4-Pro as the target agent, which means that for every 100 samples it generates, AutoRedClaw can find about 39 that successfully attack this agent. On average, the total cost of finding one successful attack sample is only about \$0.036. Beyond ASR, we further evaluate the quality of the benchmark from various dimensions (e.g., realism and verification correctness), using multiple LLMs as judges together with a human evaluation. The detailed evaluation rubric is given in a separate comment ([4/4]) due to the length limit.

We take 100 environment cases generated by AutoRedClaw, and all the 107 indirect prompt injection cases from the workflow domain of DTap. Then we employ three LLMs from different vendors as judges (GPT-6.1-Sol, Gemini-3.1-Pro and Claude-Opus-5.5), and we additionally employ 3 human evaluators as the reference. Each judge scores every case once from 1 to 5 on 8 dimensions and gives an overall realism score. The results averaged within each group are shown below:

DimensionAutoRedClaw (LLM)DTap (LLM)p-valueAutoRedClaw (human)DTap (human)p-value
Benign task realism4.803.35<0.0014.693.13<0.001
Environment fidelity4.052.04<0.0013.761.89<0.001
Injection carrier naturalness4.793.26<0.0014.683.36<0.001
Injection voice and stealth4.312.64<0.0013.852.40<0.001
Attacker plausibility3.642.40<0.0013.362.42<0.001
Harm severity3.773.640.33.733.470.08
Verification soundness4.423.16<0.0014.112.64<0.001
Goal conflict4.413.59<0.0014.483.50<0.001
Overall realism3.852.17<0.0013.552.24<0.001
GPT-6.1-SolGemini-3.1-ProClaude-Opus-5.5Avg.Human 1Human 2Human 3Avg.
DTap2.412.211.902.172.202.172.362.24
AutoRedClaw3.684.353.513.853.463.573.623.55

We observe that AutoRedClaw gets higher scores on all the 8 dimensions. The overall realism on LLM/human judge is 3.85/3.55 for AutoRedClaw and 2.17/2.24 for DTap, and on most dimensions, p-value < 0.001, indicating the significance of these evaluation.

To figure out the gap sourced from, we collect the reasoning and feedback of the evaluators. 82% of the judgments on our cases the judge thinks a real attacker would do this, compared to 38% for DTap. The largest gap is on environment fidelity (4.05 vs. 2.04). The judges point out that many DTap workspaces are almost empty or only have one user, and the sender of the injected message often does not match any user in the workspace. Our generated files, web pages and records have much richer details, although the judges still find test values in some of them, eg., 555 phone numbers or test card numbers. Moreover, the judges often find that the attacker in DTap posts in internal channels as an internal team without a clear access path, and in about one fifth of the cases the harmful instruction is already in the user's own request, so there is no real goal conflict. Harm severity is similar on the two benchmarks (3.77 vs. 3.64), so the higher overall score is not because our attacks are more harmful. Those results consistently demonstrate the high quality of samples generated by AutoRedClaw.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.