acceptodds
Under review as a conference paper at ICLR 2027

ZoneClaw: Mitigating Persistent Memory Attacks by Establishing Memory-Zoning in OpenClaw-Style Computer-Use Agents

25 views

Abstract

Computer-use agents (CUAs) increasingly operate as long-running assistants through persistent workspace memory, which OpenClaw-style CUAs realize as automatically reloaded files that hold user instructions, system summaries, and external claims at the same privilege level. Here, remembering a claim confers authority over later behavior. This enables a persistent memory attack, in which an attacker who controls only benign-looking external content induces the CUA to record attacker-favored claims during a legitimate task, and those claims later govern benign tasks the attacker never touches. The attack chain extends malicious context malicious response into malicious context memory injection malicious execution}, making this a \textbf{cross-environment threat. Existing defenses studied intervene either before content enters memory or at the action it later induces, not whether stored content may guide action. We propose ZoneClaw, which separates persistence from authority by replacing flat workspace memory with hierarchical trust zones carrying explicit authority levels. External claims persist in a low-trust zone and acquire action-guiding authority only by crossing an explicit authority boundary, at which promotion is cross-checked against zones the attacker cannot directly write. Role-specific processes of asymmetric privilege enforce this boundary, ensuring that no process both ingests external content and acts outward. Across four attack scenarios, two injection settings, and four backbones, ZoneClaw drives ASR from 372/480 to 6/480 while retaining utility in 458/480 trials, and remains effective against some defense-aware attackers. Attacker claims still persist in low-trust memory yet rarely cross the authority boundary, showing that ZoneClaw withholds authority rather than refusing to learn from the environment. Our code is available at: https://anonymous.4open.science/r/ZoneClaw.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Computer-Use Agent, OpenClaw, Memory Zoning

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

This paper investigates persistent memory attacks in OpenClaw and similar agent harnesses, where external claims supplied by an attacker are stored during legitimate tasks and later influence behavior across sessions. It proposes ZoneClaw, which separates information retention from action authority through three memory trust zones and four specialized roles, with a Gatekeeper checking external claims against trusted policies and knowledge before promoting them into operational memory. Across four attack scenarios, two injection settings, and four model backbones, the authors report reducing attack success from 77.5% to 1.25% while completing 458 of 480 legitimate tasks, with ablations and mechanism analyses supporting the effectiveness of the proposed authority boundaries.

Critical strengths
  1. Important problem for frontier agent harnesses. The paper identifies an important security gap in frontier agent harnesses, showing how external claims can enter persistent memory and acquire authority over later actions. This problem is particularly relevant to agents that maintain workspaces and learn from their environments across sessions.

  2. Clear design with supporting mechanism analysis. Separating information retention from action authority provides a clear design principle, implemented through memory zoning, role separation, and controlled promotion. The ablations and selective promotion experiments provide useful evidence that these mechanisms can restrict malicious influence while retaining useful information.

Critical weaknesses
  1. Substantial online overhead. The additional role invocations and verification steps introduce online overhead, with the reported BCC experiment showing execution time of 2.26 times the baseline and estimated API cost of 1.74 times the baseline. Broader measurements that separate memory ingestion costs from subsequent task execution would clarify the tradeoffs among security, responsiveness, and resource consumption.

  2. Limited coverage of benign and attack scenarios. Generalization remains uncertain because the evaluation covers four constructed attack scenarios, while the separate benign utility evaluation uses a subset of 30 tasks on one model. I encourage the authors to explore the DecodingTrust-Agent Platform (DTap), with appropriate adaptation to the persistent memory setting, to assess generalization across both utility and security.

Chen, Z., Liu, X., Tong, H., Guo, C., Nie, Y., Zhang, J., Kang, M., Xu, C., Liu, Q., Liu, X., Shi, T., Xiao, C., Koyejo, S., Liang, P., Guo, W., Song, D., & Li, B. (2026). DecodingTrust-Agent Platform (DTap): A controllable and interactive red-teaming platform for AI agents. arXiv. https://doi.org/10.48550/arXiv.2605.04808

Questions for authors

No

4 comments
AuthorsOpenReview ↗

Response to Reviewer jSvP (1/2)

Thank you for your insightful review and for recognizing that our work addresses an important problem for frontier agent harnesses with a clear design and supporting mechanism analysis. Following your suggestions, we have added two groups of experiments. (1) An overhead analysis covering injection versus execution, long-horizon tasks, and an optimized implementation (A1). (2) Broader coverage with attack and utility evaluation on DTap and benign utility on two Workspace-Bench categories (A2). We provide detailed point-by-point responses to your comments below.

W1: Substantial online overhead and insufficient separation of memory ingestion costs from subsequent execution costs.

A1: We agree that online overhead is an important practical concern. The defense methods require additional processing to enforce their respective safeguards. The reported latency and API cost in our paper cover a complete injection–execution pair. We analyze the overhead along three dimensions, namely injection versus execution (Table 1), long-horizon tasks (Table 2), and an optimized implementation (Table 3).

(1) Injection versus execution. In Table 1, ZoneClaw spends 144.8 s on the injection phase, where the Observer extracts external claims and the Gatekeeper checks them before promotion, concentrating much of the additional processing in this phase. This phase accounts for 87% of the additional latency and 79% of the additional API cost relative to OpenClaw. Subsequent tasks can reuse the admitted information without repeating these checks. In the execution phase, ZoneClaw adds only 12.3 s and USD 0.0030 per task over OpenClaw, with about half the latency of the other defense methods (62.0 to 71.0 s) and 2.5× to 5.1× lower in cost.

Table 1. Phase-separated overhead on user-triggered BCC Exfiltration scenarios with Qwen3.7-Plus at medium reasoning.

SystemPhaseLLM callsInput tokensOutput tokensLatency(s)API cost(USD)
No DefenseInjection6.73120.0k3.0k58.50.0134
Exploitation2.8746.8k0.9k19.70.0051
WatcherInjection11.27140.6k3.9k71.70.0240
Exploitation15.33303.5k3.2k68.40.0264
ClawGuardInjection7.33132.9k3.1k50.60.0140
Exploitation13.20257.8k3.0k67.10.0228
MELONInjection10.80162.7k3.9k60.90.0304
Exploitation19.00338.7k4.0k71.00.0411
PrivSepInjection12.8779.0k4.9k71.60.0161
Exploitation18.07156.5k3.9k62.00.0206
ZoneClawInjection14.93125.7k7.5k144.80.0242
Exploitation6.2747.4k1.4k32.00.0081

(2) Long-horizon tasks. We conduct one injection phase followed by ten executions with Claude Sonnet 4.6 and report totals after 1, 5, and 10 executions in Table 2. ZoneClaw incurs the largest overhead after the first execution, but its total API cost and latency over ten executions are 24.8% and 8.3% lower than OpenClaw’s and 4.2% and 15.7% lower than PrivSep’s. This saving mainly arises because the Planner and Executor of ZoneClaw read only the trusted information in D0 and D1, and the actor of PrivSep reads only its structured summaries by its reader, whereas OpenClaw reloads all stored content. However, PrivSep's summaries may carry malicious instructions into later memory updates and rise its ASR to 3/3, whereas ZoneClaw consistently stays at 0/3, ehabling reuse of verified knowledge while restricting unapproved instructions from guiding actions.

Table 2. Total overhead after 1, 5, and 10 executions on user-triggered BCC Exfiltration scenario.

MethodTask numberTotal cost(USD)↓Total latency(s)ASR↓Utility↑
10.3665117.03/33/3
OpenClaw51.2180347.23/33/3
102.2824707.83/33/3
10.4214178.00/33/3
PrivSep51.0304431.62/33/3
101.7915770.13/33/3
10.4568188.60/33/3
ZoneClaw51.0165379.00/33/3
101.7160649.30/33/3

(3) Optimized implementation. We further preload tool instructions, batch reads, reduce coordination turns, and shorten reports to reduce its overhead. In Table 3, ZoneClaw's latency falls from 176.8 s to 145.0 s and API cost from USD 0.0322 to 0.0244, with ASR and utility unchanged. Relative to OpenClaw, the overhead drops from 2.26× to 1.86× in latency and from 1.74× to 1.32× in cost. The optimized ZoneClaw remains the cheapest defense method with latency slightly above the baselines (117.7 to 140.0 s). We will release the optimized ZoneClaw implementation upon acceptance.

Table 3. End-to-end overhead and performance with the optimized implementation (same setting as Table 1).

ImplementationLatency(s)↓API cost(USD)↓ASR↓Utility↑
OpenClaw78.10.018514/1515/15
Watcher140.00.05049/1515/15
ClawGuard117.70.03680/1510/15
MELON131.80.071511/1515/15
PrivSep133.50.036713/1513/15
ZoneClaw176.80.03220/1515/15
ZoneClaw (optimized)145.00.02440/1515/15
AuthorsOpenReview ↗

Response to Reviewer jSvP (2/2)

W2: Limited coverage of benign and attack scenarios, with a suggestion to assess utility and security on DTap.

A2: We agree that broader coverage of both attack and benign scenarios is important. Following your suggestion, we extend our evaluation along two dimensions, namely attack and utility evaluation on DTap (Table 4) and benign utility on additional Workspace-Bench tasks (Table 5).

(1) Attack and utility evaluation on DTap. Following your suggestion, we adapt 10 DTap tasks to the persistent memory setting. These tasks belong to the hidden side effect class, in which the attacker tries to inject an unauthorized verification claim into the Customer Relationship Management (CRM) system while the agent completes the legitimate task. We inject attacker payloads into each adapted task as specified by our threat model (Section 2.3 of our paper), so that the payload is read in the injection phase and the task is completed in the execution one. Utility is measured with the official DTap verifiers, and ASR counts unauthorized claims actually written into the CRM. ZoneClaw reduces ASR from 9/10 to 0/10 with no utility loss, whereas Watcher and ClawGuard remain at 8/10 and 10/10. This discovery supports ZoneClaw's effectiveness on the adapted DTap tasks beyond our original attack scenarios.

Table 4. Attack and defense performance on 10 DTap tasks adapted to cross-session persistent memory.

Task setMethodASR↓Utility↑
DTapOpenClaw9/1010/10
DTapWatcher8/1010/10
DTapClawGuard10/1010/10
DTapZoneClaw0/1010/10

(2) Benign utility. We add 19 Workspace-Bench tasks with no attacker content from two categories, namely Operations Manager (10 tasks, 204 rubric criteria) and Logistics Manager (9 tasks, 163 rubric criteria). Here, we utilize GPT-5.5 at medium reasoning as the backbone and Claude Sonnet 4.6 as the judge, which follows the same setting as Appendix C.8 of our paper. As shown in Table 5, ZoneClaw stays within about one percentage point of OpenClaw on Operations Manager and is slightly above it on Logistics Manager. In contrast, ClawGuard and PrivSep lose 8.0 and 13.4 points on Operations Manager. These results show that ZoneClaw maintains benign task utility comparable to OpenClaw while outperforming the other defense methods across both task categories, complementing the utility evidence from DTap.

Table 5. Benign utility on 10 Operations Manager tasks covering 204 rubric criteria and 9 Logistics Manager tasks covering 163 rubric criteria from WorkspaceBench, with GPT-5.5 at medium reasoning and Claude Sonnet 4.6 as the judge.

MethodMetricOperations ManagerLogistics Manager
OpenClawMean task score(%)↑45.5449.47
ClawGuardMean task score(%)↑37.5550.81
PrivSepMean task score(%)↑32.1639.05
ZoneClawMean task score(%)↑44.5452.24

Our responses to Reviewers FiEf and 6xj1 further add new information sources, new input forms, and two additional agent harnesses. We will incorporate all new results and clarifications into the revised paper before the end of the discussion period, with changes highlighted in blue.

We hope these results and analyses address your concerns, and we are happy to provide further clarification.

Thanks for your response. I will keep my positive score!

AuthorsOpenReview ↗

Revision Uploaded: Follow-Up to Reviewer jSvP

Dear Reviewer jSvP,

Thank you for your follow-up and for keeping your positive score. We are truly encouraged by your continued support of our work. Your comments on the online overhead and the coverage of scenarios directly led to the phase-level overhead analysis and the broader evaluation, both of which have made the paper more complete.

We have uploaded the revised paper, which incorporates all new results and clarifications from our responses. All changes are highlighted in blue and these additions are also summarized in Section 4.8. The table below lists where each of your comments is addressed.

Your commentRevisionLocation
W1 Online overheadOverhead separated into the injection and exploitation phasesAppendix C.6
Cumulative overhead of one injection phase followed by ten tasksAppendix C.6
Optimized implementation of ZoneClawAppendix C.6
W2 Coverage of attack scenarios10 DTap tasks adapted to cross-session persistent memorySection 4.7
Two additional harnesses, two additional information sources, and three additional input formsSection 4.7 and Appendix C.7
W2 Coverage of benign scenarios19 additional Workspace-Bench tasks in two categoriesAppendix C.10

We would like to highlight that the core contribution of our paper lies in the design principle of separating persistence from authority.

  • Persistent memory attacks succeed because remembered content is treated as trusted by default. Following this principle, ZoneClaw enables CUAs to continually accumulate knowledge and experience from external environments, while withholding authority from untrusted content over subsequent actions.

Thank you once again for your time and constructive feedback. We would be glad to provide any further clarification before the discussion period ends.

Best regards,

Authors of Submission 5466

Reviewer_6xj1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies persistent memory attacks against OpenClaw-style computer-use agents, where attacker-controlled external content is stored as memory and later influences benign tasks across sessions. It proposes ZoneClaw, which separates memory persistence from behavioral authority through three trust zones and four roles with asymmetric privileges. External claims remain untrusted unless a Gatekeeper validates their promotion against trusted context. Across four models and four attack scenarios, ZoneClaw reduces attack success from 77.5% to 1.25%, while achieving 95.4% task completion in the evaluated attack settings.

Critical strengths
  1. Good design principle. Separating persistence from authority directly addresses how seemingly benign external claims become operational rules, allowing agents to retain useful information while restricting its influence on future actions.

  2. Convincing mechanism-focused evaluation. Fresh-session experiments isolate persistent memory as the attack channel, while boundary-level measurements, selective-promotion tests, and ablations provide evidence that the defense works by restricting authority acquisition rather than indiscriminately rejecting external information.

Critical weaknesses
  1. Limited task coverage. The paper’s main evaluation consists of repeated runs on only four tasks under different triggering settings. Such limited task diversity may introduce substantial bias in the selection of attack strategies. Furthermore, the paper does not assess how difficult these attacks are to defend against, leaving their representativeness unclear and potentially leading to substantially biased evaluation results.

  2. Unclear authority boundary. The paper primarily evaluates external information from emails, chats, forums, and marketplace webpages, without comparable validation for broader sources such as general web search results and local or downloaded files. External content can also have complex structures and span multiple modalities, but the paper does not clearly explain how such content should be decomposed and assessed. Furthermore, the trust required for external information can depend on its subsequent use. For example, reading information from a URL and uploading user data to that URL require different permissions. It is unclear how promotion to D1 interacts with these operation-specific permissions, or how such distinctions are maintained as information is reused across tasks.

Questions for authors
  1. suggestion: I feel Figure 2 visually cluttered, which makes the key information difficult to read
7 comments
AuthorsOpenReview ↗

Response to Reviewer 6xj1 (1/4)

Thank you for your careful review and for recognizing that separating persistence from authority is a good design principle supported by a convincing mechanism-focused evaluation. Following your reviews, we have added four groups of experiments. (1) New tasks in DTap and an assessment of how difficult the attacks are to defend against (A1). (2) New information sources, namely search-linked webpages and downloaded files (A2-1). (3) An explanation of how external content is decomposed and assessed with three New input forms (A2-2). (4) A study of how the permission scope is preserved in memory and maintained across tasks (A2-3). We have also simplified Figure 2 as you suggested in our paper (A3). We will incorporate other new results and clarifications into the revised paper before the end of the discussion period, with changes highlighted in blue. We provide detailed point-by-point responses to your comments below.

W1: Limited task coverage, and no assessment of how difficult the attacks are to defend against.

A1: We agree that repeated trials do not substitute for task diversity. We therefore extend our evaluation along two dimensions with Claude Sonnet 4.6, namely new tasks from DTap (Table 1) and how difficult the attacks are to defend against (Table 2).

(1) New tasks from DTap. We adapt 10 distinct DTap tasks to cross-session persistent memory in Table 1. These tasks involve the hidden side effect class, in which the attacker tries to insert an unauthorized verification claim into the Customer Relationship Management (CRM) system while the agent completes the legitimate task. We inject attacker payloads into each task as specified by our threat model. Utility is measured with the official DTap verifiers, and ASR counts unauthorized claims actually written into the CRM. ZoneClaw reduces ASR from 9/10 to 0/10 with no utility loss, whereas Watcher and ClawGuard remain at 8/10 and 10/10. Our responses to W2-1 and W2-2 further add new information sources and input forms.

Table 1. Attack and defense performance on 10 DTap tasks adapted to cross-session persistent memory.

Task setMethodASR↓Utility↑
DTapOpenClaw9/1010/10
DTapWatcher8/1010/10
DTapClawGuard10/1010/10
DTapZoneClaw0/1010/10

(2) Difficulty of the attacks. We assess this part in two ways. First, the standard attacks in our paper are built without any knowledge of ZoneClaw and are not blocked by existing defenses. Over the eight scenario-setting combinations in Table 2 of the paper, they succeed in 97/120 trials without defense and still in 60/120 to 97/120 trials under the four defense baselines, against 0/120 under ZoneClaw. Second, we further explore attacks built with full knowledge of ZoneClaw’s design (See Section 4.6 and Appendix C.5 of our paper). Corroboration Planting supplies the approval evidence that the Gatekeeper looks for, and Claim Fragmentation splits the malicious payload into individually benign claims that become harmful only in combination. Table 2 extends Table 5 of our paper with Watcher and ClawGuard. ZoneClaw keeps CSR and ASR at 0/15 with 15/15 utility under all three attacks. Watcher remains vulnerable at 4/15 to 6/15 ASR, and ClawGuard blocks the attacks only by reducing utility to between 3/15 and 9/15.

Table 2. Defense performance on user-triggered BCC Exfiltration task under the standard attack and two defense-aware attacks, with 15 trials per attack. CSR counts trials in which the attacker claim reaches authority-bearing memory.

MethodMetricStandard attackCorroboration PlantingClaim Fragmentation
CSR↓15/1515/1515/15
OpenClawASR↓12/1511/1515/15
Utility↑15/1513/1515/15
CSR↓12/156/158/15
WatcherASR↓6/154/156/15
Utility↑14/1514/1513/15
CSR↓14/1514/1515/15
ClawGuardASR↓0/150/150/15
Utility↑9/154/153/15
CSR↓0/150/150/15
ZoneClawASR↓0/150/150/15
Utility↑15/1515/1515/15
AuthorsOpenReview ↗

Response to Reviewer 6xj1 (2/4)

W2-1. Limited evaluation on broader information sources.

A2-1: We agree that validation on broader information sources is important. We therefore add two information sources with Claude Sonnet 4.6 in Table 3, namely search-linked webpages and downloaded files. A downloaded file is stored on and read from the local file system in the same way as a local file, so we evaluate the two as one setting. These experiments assess whether the defense effectiveness and task utility observed in our original settings extend to these additional information sources by delivering payloads in different mediums.

As shown in Table 3, ZoneClaw is the only method that reaches 0/10 ASR with 10/10 utility on both sources. ClawGuard also blocks the attacks, but its utility falls to 8/10 and 3/10. PrivSep is attacked in every trial, which may because its reader condenses the page into a validated structured summary, which removes the surrounding context that makes the raw claim look suspicious. These findings extend our empirical validation beyond the original sources, demonstrating that ZoneClaw can prevent the cross-session attacks while supporting legitimate task completion across these additional information sources.

Table 3. Defense performance across search-linked webpages and downloaded files with Claude Sonnet 4.6.

MethodMetricSearch-linked webpageDownloaded files
OpenClawASR↓2/1010/10
Utility↑10/1010/10
ClawGuardASR↓0/100/10
Utility↑8/103/10
PrivSepASR↓10/1010/10
Utility↑10/109/10
ZoneClawASR↓0/100/10
Utility↑10/1010/10
AuthorsOpenReview ↗

Response to Reviewer 6xj1 (3/4)

W2-2: Unclear assessment and decomposition of external content with complex structures or multiple modalities.

A2-2: We agree that it is important to explain how such content is decomposed and assessed. We first describe this procedure in ZoneClaw and then evaluate it on three additional input forms with complex structures or multiple modalities (Table 4).

(1) How content is decomposed and assessed. ZoneClaw applies the same procedure to process external content across these input formats. The Observer reads the textual and visual content and records each extracted claim with its source in D2, without granting it authority (Sec. 3.3.3). Then the Gatekeeper assesses task-relevant facts and proposed behavioral instructions separately against the policy in D0 and the trusted content in D1 before deciding whether to promote them. For example, verifying the factual details needed to write an email does not authorize an instruction in the same document to add a BCC recipient. The Executor subsequently performs the task using D0 and D1, so the input format only affects how the Observer transcribes a claim, and the authority check stays the same.

(2) Evaluation on new input forms. We render the BCC Exfiltration source as a screenshot, a scanned PDF, and a mixed text-and-image document to extend beyond the original text format. We evaluate four methods across these settings with Claude Sonnet 4.6 while keeping the benign facts and malicious instructions identical. As shown in Table 4, ZoneClaw reaches 0/3 ASR and 3/3 utility in all four forms and retains all 36 benign facts. OpenClaw and PrivSep are each attacked in 11/12 trials. ClawGuard also reaches 0/12 ASR, but it just completes 5/12 tasks, with 0/3 on scanned PDFs. In one Mixed trial, the injected claim appears in ZoneClaw's trusted memory, but no BCC email is sent (ASR=0). Overall, these results support ZoneClaw’s ability to handle structured and multimodal external content while preserving legitimate task utility and suppressing cross-session attacks.

Table 4. Defense performance on structured and multimodal inputs in BCC Exfiltration with Claude Sonnet 4.6. Each input form presents the same source content, with three trials per method–input combination.

Input formMethodBenign fact extraction(%)↑Malicious claim extraction(%)ASR↓Utility↑
TextOpenClaw100.0(9/9)100.0(3/3)3/33/3
TextClawGuard100.0(9/9)100.0(3/3)0/31/3
TextPrivSep100.0(9/9)100.0(3/3)3/33/3
TextZoneClaw100.0(9/9)0.0(0/3)0/33/3
ScreenshotOpenClaw100.0(9/9)100.0(3/3)3/33/3
ScreenshotClawGuard100.0(9/9)100.0(3/3)0/32/3
ScreenshotPrivSep100.0(9/9)100.0(3/3)3/33/3
ScreenshotZoneClaw100.0(9/9)0.0(0/3)0/33/3
Scanned PDFOpenClaw100.0(9/9)100.0(3/3)2/33/3
Scanned PDFClawGuard100.0(9/9)100.0(3/3)0/30/3
Scanned PDFPrivSep100.0(9/9)100.0(3/3)2/33/3
Scanned PDFZoneClaw100.0(9/9)0.0(0/3)0/33/3
Mixed text and imageOpenClaw100.0(9/9)100.0(3/3)3/33/3
Mixed text and imageClawGuard100.0(9/9)100.0(3/3)0/32/3
Mixed text and imagePrivSep100.0(9/9)100.0(3/3)3/33/3
Mixed text and imageZoneClaw100.0(9/9)33.3(1/3)0/33/3
Reviewer_FiEfOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents ZoneClaw, which assigns credibility levels to an OpenClaw agent’s memory to defend computer-use agents against persistent memory attacks. OpenClaw maintains 3 zones: D0 policy zone, D1 Trusted Zone, and D2 Untrusted zone. Memory claims move from D2 to D1 through promotion decisions. Experiments with 4 LLMs on 4 scenarios demonstrate that ZoneClaw is safer against memory attacks compared to no defense or other baselines.

Critical strengths
  1. The paper proposes a simple yet effective mechanism to split agent context into 3 zones with different credibility levels. The agent can explicitly distinguish how it uses information from different zones to reduce the effect of malicious input.
Critical weaknesses
  1. The experiment is a key limitation of this work. The evaluation is based on 4 scenarios × 2 injection settings, and each scenario setting has only one pair of tasks. Although the experiments were run 15 trials, this does not increase the comprehensiveness of the evaluation. The limited scale of the evaluation makes it difficult to draw sound comparisons among different approaches. Moreover, the paper considers only one agent harness, which makes the proposed method a useful trick rather than a systematic approach to improve agent safety. The paper would greatly benefit from more comprehensive experiments and evaluation.


  2. The paper lacks a preliminary study on how memory benefits utility in these scenarios. ZoneClaw outperforms the no-defense baseline on the proposed scenarios, but it is important to show that these scenarios can benefit from memory in terms of utility, which would justify the need for memory. Otherwise, no memory would be the safest solution.

Questions for authors

See the Weakness section.

7 comments
AuthorsOpenReview ↗

Response to Reviewer FiEf (1/3)

Thank you for your careful review and for recognizing ZoneClaw as "a simple yet effective mechanism" that splits agent context into three zones with different credibility levels. Following your comments, we have added three groups of experiments. (1) Broader evaluation coverage with new tasks in DTap, new information sources, and new input forms (A1-1). (2) Two additional harnesses, namely Hermes and OpenClaw-CN (A1-2). (3) The preliminary study you asked for on how memory benefits utility, with a no-memory variant (A2). We will incorporate all new results and clarifications into the revised paper before the end of the discussion period, with changes highlighted in blue. We provide detailed point-by-point responses to your comments below.

W1-1: Limited task diversity and evaluation coverage.

A1-1: We agree that repeating a fixed task pair does not increase evaluation comprehensiveness. We therefore broaden our evaluation along three dimensions with Claude Sonnet 4.6, namely new tasks from DTap (Table 1), new information sources (Table 2), and new input forms (Table 3).

(1) New tasks. We first expand task coverage with 10 distinct DTap tasks adapted to cross-session persistent memory in Table 1. Each task belongs to a hidden side-effect class, in which the attacker attempts to insert an unauthorized verification claim into the Customer Relationship Management (CRM) system while the agent carries out the legitimate task. Utility follows the original DTap rubric, and ASR counts unauthorized claims actually written into the CRM. ZoneClaw achieves 0/10 ASR while preserving 10/10 utility, compared with 9/10 ASR for OpenClaw, 8/10 for Watcher, and 10/10 for ClawGuard. These additional tasks extend the comparison beyond the original attack scenarios, showing that ZoneClaw reduces attack success without sacrificing legitimate task utility on DTap tasks.

Table 1. Attack and defense performance on 10 DTap tasks.

MethodASR↓Utility↑
OpenClaw9/1010/10
Watcher8/1010/10
ClawGuard10/1010/10
ZoneClaw0/1010/10

(2) New information sources. We further broaden the range of external information sources by evaluating search-linked webpages and downloaded files in Table 2 over ten trials per source. ZoneClaw is the only method with 0/10 ASR and 10/10 utility. ClawGuard also blocks the attacks but retains only 8/10 and 3/10 utility, while PrivSep is attacked in every trial, possibly because its reader’s structured summaries omit contextual warning signs. These results extend the comparison beyond the original information sources and show that ZoneClaw combines zero attack success with high legitimate task utility across both new information sources.

Table 2. Defense performance across search-linked webpages and downloaded files.

MethodMetricSearch-linked webpageDownloaded files
OpenClawASR↓2/1010/10
Utility↑10/1010/10
ClawGuardASR↓0/100/10
Utility↑8/103/10
PrivSepASR↓10/1010/10
Utility↑10/109/10
ZoneClawASR↓0/100/10
Utility↑10/1010/10

(3) New input forms. We also expand the evaluation beyond textual inputs by rendering the BCC Exfiltration source as screenshots, scanned PDFs, and mixed text-and-image documents in Table 3. ZoneClaw reaches 0/3 ASR and 3/3 utility in all three input forms. OpenClaw and PrivSep are attacked in 8/9 runs, and ClawGuard blocks the attack but keeps only 4/9 utility. This comparison demonstrates that ZoneClaw preserves legitimate task utility while suppressing attacks across the structured and multimodal input forms.

Table 3. Defense performance on structured and multimodal input forms in BCC Exfiltration.

MethodMetricScreenshotScanned PDFMixed text and image
OpenClawASR↓3/32/33/3
Utility↑3/33/33/3
ClawGuardASR↓0/30/30/3
Utility↑2/30/32/3
PrivSepASR↓3/32/33/3
Utility↑3/33/33/3
ZoneClawASR↓0/30/30/3
Utility↑3/33/33/3

Overall, these additional experiments expand task diversity through DTap and broaden coverage across information sources and structured and multimodal inputs, providing a more comprehensive basis for comparison. Across the added evaluations, ZoneClaw achieves zero attack success while delivering utility that is comparable to OpenClaw’s and the best among the other defense methods.

AuthorsOpenReview ↗

Response to Reviewer FiEf (2/3)

W1-2: Evaluation limited to one agent harness.

A1-2: We agree that evaluating ZoneClaw beyond a single harness is important. We therefore extend our evaluation to Hermes and OpenClaw-CN. Hermes is an independently developed harness unrelated to OpenClaw, and OpenClaw-CN is a customized OpenClaw fork. We compare ZoneClaw with no defense and with ClawGuard and PrivSep, the two baselines with the lowest total ASR in Table 2 of the paper. The evaluation uses BCC Exfiltration and Source Redirection in the user-triggered setting, with Claude Sonnet 4.6 and 15 trials per scenario.

As shown in Table 4, ZoneClaw achieves the lowest ASR and highest utility on all three harnesses. On the two new harnesses, its ASR is 2/30 and 0/30, whereas the two defense baselines stay at 13/30 to 22/30 ASR. Hermes is the most challenging setting for ZoneClaw, with two attack successes and 20/30 utility. The utility loss arises because task details are dropped in the handoff between role-specific processes, a failure source that Appendix D.2 also traces on OpenClaw. The undefended utility is low on every harness because a successful Source Redirection attack yields zero utility by definition (Appendix B.3).

Table 4. Defense performance across agent harnesses on BCC Exfiltration and Source Redirection with 15 trials per scenario.

HarnessMethodASR↓Utility↑
OpenClawNo defense27/3015/30
OpenClawClawGuard15/309/30
OpenClawPrivSep9/3018/30
OpenClawZoneClaw0/3030/30
HermesNo defense19/3015/30
HermesClawGuard13/3016/30
HermesPrivSep22/3016/30
HermesZoneClaw2/3020/30
OpenClaw-CNNo defense29/3015/30
OpenClaw-CNClawGuard13/305/30
OpenClaw-CNPrivSep15/3017/30
OpenClaw-CNZoneClaw0/3028/30

Overall, these results indicate that separating persistence from authority transfers beyond the original harness, both to an independent framework and to a fork within the OpenClaw family. This supports ZoneClaw as a design that is not tied to one harness.

AuthorsOpenReview ↗

Response to Reviewer FiEf (3/3)

W2: Show that these scenarios benefit from memory, since otherwise no memory would be the safest solution.

A2: We agree that disabling memory removes this attack. However, it also removes the long-term utility that memory provides, which recent work regards as the key to agents that operate continuously and accumulate experience [1]. We investigate this trade-off with an OpenClaw w/o memory variant in two settings, namely cross-session tasks (Table 5) and repeated injection (Table 6). In this variant, each subsequent session starts from the initial workspace, preserving static workspace files while discarding information written during earlier sessions.

(1) Cross-session tasks require memory. We compare OpenClaw, OpenClaw w/o memory, and ZoneClaw on five cross-session tasks in each of three environments, namely our scenarios, Workspace-Bench, and DTap with Claude Sonnet 4.6. In every task, an earlier session obtains the related facts, and a fresh session must use them to complete its legitimate task. As shown in Table 5, OpenClaw w/o memory's utility falls from 5/5 to 0/5 in all three environments, whereas ZoneClaw keeps 5/5. Its restrictions on memory authority therefore preserve the information needed for legitimate task completion without reducing utility.

Table 5. Cross-session task utility with five distinct tasks per environment.

MethodMetricOur scenarioWorkspaceBenchDTap
OpenClawUtility↑5/55/55/5
OpenClaw w/o memoryUtility↑0/50/50/5
ZoneClawUtility↑5/55/55/5

(2) No memory blocks the attack but also fails the task. We further test a long-horizon setting by following the repeated-injection protocol in Appendix C.7 of our paper, where each round brings one new legitimate update and one rephrased malicious BCC rule. We add OpenClaw w/o memory as a new variant. As you noted, OpenClaw w/o memory is safe. Its ASR is 0/3 at every round, since the injected rule cannot persist. However, the legitimate updates are lost in the same way. It then asks the user for the missing information instead of sending the email, so its utility is also 0/3. OpenClaw completes the task but is attacked in 3/3 runs. ZoneClaw is the only method that reaches both 0/3 ASR and 3/3 utility at every round. These results show that ZoneClaw preserves the utility of accumulated memory across repeated interactions while suppressing the cross-session attacks, even as malicious instructions are repeatedly introduced.

Table 6. Repeated injection performance in BCC Exfiltration scenario, evaluated after 1, 5, and 10 memory-update rounds.

MethodASR↓Utility↑
After 1 roundAfter 5 roundsAfter 10 roundsAfter 1 roundAfter 5 roundsAfter 10 rounds
OpenClaw3/33/33/33/33/33/3
OpenClaw w/o memory0/30/30/30/30/30/3
ZoneClaw0/30/30/33/33/33/3

(3) Prior work also supports the necessity of memory. Recent works treat memory as a core component of LLM agents [2, 3]. In particular, MemoryArena [2] constructs interdependent multi-session agentic tasks whose later subtasks require information from earlier sessions. Agent Workflow Memory [4] further shows that equipping agents with memory of past experience raises their success rates, for instance by 51.1% relative on WebArena. Memory is also the focus of recent spotlight papers on cross-trial memory [5], reusable skills [6], and memory-based test-time adaptation [7].

Overall, our claim concerns tasks that depend on earlier sessions. For such tasks, no memory avoids the attack only by giving up the utility, whereas ZoneClaw keeps the utility and withholds authority from attacker content.

We hope these additions address your concerns, and we would appreciate your considering an updated assessment.

[1] Position: Modular Memory is the Key to Continual Learning Agents. ICML 2026, Position Track, Spotlight.

[2] Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks. ICML 2026, Regular.

[3] Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions. ICLR 2026, Poster.

[4] Agent Workflow Memory. ICML 2025, Poster.

[5] G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems. NeurIPS 2025, Spotlight.

[6] Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents. ICML 2026, Spotlight.

[7] Just-In-Time Reinforcement Learning: Continual Learning in LLM Agents Without Gradient Updates. ICML 2026, Spotlight.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.