acceptodds
Under review as a conference paper at ICLR 2027

ELLA: Weaponizing Visual Evidence to Hijack Memory Formation in Self-Evolving Agents

Abstract

Persistent memory enables self-evolving agents to summarize interactions and revise state across tasks, but it also allows untrusted evidence to be combined with the agent’s own behavior and persisted as trusted memory. Existing memory-poisoning attacks place a malicious instruction, trigger, or target relation in attacker-controlled content before memory writing. We expose a different attack surface: an attacker can provide only source-side visual evidence and induce the victim’s own memory writer to complete a harmful source–action association during an otherwise ordinary update. We introduce ELLA, which selects benign-looking source evidence that is likely to be retrieved and consumed while leaving the target action and final relation absent from the submitted content. Across multimodal benchmarks, three memory backends, and multiple reader models, ELLA induces persistent targeted behavior after all attacker-owned sources are removed.A no-completion control eliminates the effect despite full source retrieval, and a commitment-flip experiment shows that the activated action follows evidence supplied after the carrier is fixed. These results identify memory formation itself as a security boundary: harmful persistent state can be created by the trusted writer rather than directly injected by the attacker.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Self-Evolving Agents, Multimodal Agents, Memory Poisoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7Qv1OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates an interesting security problem in self-evolving agents: untrusted external evidence can be combined with the agent's own actions and feedback to form persistent behavioral associations during memory updates. The authors propose ELLA, which uses Semantic Anchoring (SA) and Writer Exposure (WE) to select visual evidence that is likely to be retrieved and incorporated into the agent's memory. Unlike traditional memory poisoning attacks, the target action is not explicitly included in the injected content but is instead associated with the source evidence through the agent's own memory writer. Experiments across multiple benchmarks, memory backends, and model backbones demonstrate the effectiveness of the proposed attack.

Critical strengths
  • Interesting and important research problem. The paper identifies an interesting attack surface in the memory formation process of self-evolving agents. I particularly appreciate the perspective that harmful persistent associations can be generated by the agent's own memory writer, rather than being directly injected by the attacker.

  • Clear writing and motivation. The paper is generally well written, and the motivation and overall attack pipeline are clearly presented.

  • Interesting mechanism experiments. I appreciate the causal experiments, particularly the no-completion control, source removal, lineage cleanup, and commitment-flip experiments. These provide useful evidence that the observed attack effects are associated with writer-generated memory rather than merely the retrieval of malicious content.

Critical weaknesses
  • Why visual evidence? The proposed method is straightforward and easy to understand. However, I wonder whether the vulnerability is actually specific to visual evidence. For example, what if we replace the image with an equivalent textual description? Would the attack still succeed?

  • Limited exploration of image optimization. The proposed method mainly relies on selecting existing images through semantic similarity and UCB-based feedback. However, it remains unclear why image selection is preferred over directly optimizing or generating images for the attack objective. Could image optimization lead to stronger or more transferable attacks? Exploring this alternative would help better justify the current design and potentially strengthen the methodological contribution.

  • Insufficient understanding of source–action associations. One interesting question is why the memory writer associates the source evidence with a particular target action, especially when the target action is absent from the injected content. I wonder how much the attack effectiveness depends on the semantic relationship between the source and the target action. For example, would the attack still succeed when the two are weakly related or unrelated? A more systematic analysis could provide valuable insights into the underlying memory formation mechanism.

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

W1. What changes when an original image is replaced by a text description of the same scene?

We sincerely appreciate your positive assessment of the memory-formation perspective and your recognition of the paper's motivation, presentation, and causal analyses. We also thank you for the constructive suggestions regarding input modality, direct image optimization, and semantic relatedness. Following these suggestions, we have conducted paired image–caption and image-optimization comparisons, together with a controlled semantic-relatedness analysis.

We appreciate the reviewer's suggestion to test whether equivalent textual evidence can induce the same mechanism. We first compare original images with captions describing the same scenes on 100 matched queries. The image group had a 14 pp higher source retrieval rate and a 55 pp higher rate of source exposure to the Writer. We observe that the advantage of visual evidence becomes considerably larger at the memory-writing boundary, not merely at initial retrieval.

Table 7. Paired image–caption retrieval and Writer exposure

Input conditionQueriesCorresponding-source retrieval rateWriter source exposure
Original image10093%81%
Caption describing the same scene10079%26%
Image − caption100 pairs+14 pp+55 pp

We separately ran an end-to-end evaluation with Qwen3-VL-30B-A3B + Mem0 + Mem-Gallery. All four conditions shared the same 71 queries on which the paired Clean run did not select the target action. SC-ASR additionally requires retrieval of a traceable derived memory.

Table 8. End-to-end comparison of image and caption conditions

Input conditionConditional Post-ASRConditional SC-ASR
Visual PFS19/71 (26.8%)11/71 (15.5%)
Visual Static21/71 (29.6%)8/71 (11.3%)
Caption PFS42/71 (59.2%)0/71 (0.0%)
Caption Static43/71 (60.6%)0/71 (0.0%)

We acknowledge that the caption conditions show higher overall target-action rates. However, we believe the more informative distinction for ELLA is whether the target action is accompanied by traceable, writer-generated memory. In the separate end-to-end evaluation, the Caption PFS source did not appear in the retrieval results, and none of the 27 derived memories linked back to it. The 26% caption Writer exposure in Table 7 belongs to a distinct 100-query paired-retrieval protocol. In the end-to-end experiment, only the visual conditions yielded target-action hits accompanied by retrieval of source-linked derived memories. We will clarify these protocol differences and the lineage criteria in the revised paper.

AuthorsOpenReview ↗

W2. Why study selection from existing images, and can direct optimization improve success rates or cross-model transfer?

We appreciate the reviewer's suggestion to directly compare image selection with optimization. We believe image selection is a meaningful methodological choice because ELLA investigates whether persistent memory poisoning can arise from ordinary, unmodified visual evidence, without attacker-controlled pixel-level perturbations. ELLA studies a setting in which an external provider can choose and submit an existing, unmodified image. Even with this limited control over the input, we observed that the system’s own Writer can form persistent, traceable associations.

To further assess the reviewer's concern, we conduct a paired comparison of existing image selection and direct image optimization. Both image sets were determined on Qwen and then held fixed for evaluation on an independent DeepSeek model. Within each model, both methods used the same eligible queries.

Table 9. Paired comparison of existing image selection and direct image optimization

Evaluation modelShared eligible queriesSelection Post-ASROptimization Post-ASRPost-ASR differenceSelection SC-ASROptimization SC-ASRSC-ASR difference
Qwen3-VL-30B-A3B499/49 (18.4%)5/49 (10.2%)−8.2 pp3/49 (6.1%)1/49 (2.0%)−4.1 pp
DeepSeek-V4.1-Flash6744/67 (65.7%)44/67 (65.7%)0.0 pp33/67 (49.3%)33/67 (49.3%)0.0 pp

Differences are calculated as optimization minus selection.

We observe no improvement from direct optimization in either evaluated model: selection attains higher point estimates on Qwen and has the same point estimates as optimization on DeepSeek for both metrics. We believe the comparison strengthens our design rationale: selecting unmodified images is sufficient to induce traceable persistent-memory effects, and pixel-level optimization is not necessary for the vulnerability to manifest in these settings. The existing-image selection condition also produces traceable hits on the independent model.

AuthorsOpenReview ↗

W3. How does semantic relatedness between the source and target action affect memory formation?

We appreciate the reviewer's request for a more systematic analysis of source–target semantic relatedness. We therefore hold the carrier images, target action, Writer, backend, and random seeds fixed and vary only the semantic relationship between the supplemental source text and the target action. The high-relatedness condition explicitly mentioned the target action and served as a diagnostic control. The unrelated, low-relatedness, and medium-relatedness conditions did not explicitly instruct the agent to take the target action.

Table 10. Semantic relatedness and downstream outcomes

Semantic relatednessRule-based scoreAssociation formation ratePost-ASRPost-ASR difference vs. unrelated
High588.0%53.8%+17.9 pp
Medium283.0%48.7%+12.8 pp
Low174.0%48.7%+12.8 pp
Unrelated060.0%35.9%Reference

We observe two complementary findings. First, semantic relatedness facilitates writer-side association formation: among conditions without explicit target-action instructions, the formation rate rises from 60.0% (Unrelated) to 74.0% (Low) and 83.0% (Medium). The Medium and Low conditions have Post-ASR values 12.8 pp above Unrelated. Second, we believe this result also underscores the breadth of the underlying risk: even unrelated source context yields a 60.0% observed association-formation rate. Thus, strong semantic alignment improves formation but is not a strict prerequisite. The High condition explicitly mentions the target action and is treated only as a diagnostic control.

We sincerely appreciate the reviewer's constructive suggestions. We hope these additional analyses clarify why visual evidence and selection-based optimization are meaningful for ELLA, and encourage you to reconsider your assessment.

Thanks for the authors' reply. I appreciate the experiments to strengthen the method design. However, the current rebuttal shows the performance but not the intuition. I would like the authors to discuss more insights behind those experimental results. For example, why can selection outperform generation? and the same for the following questions. I would have been more willing to champion it with the insights behind the work.

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

This paper proposes ELLA, which shows how benign-looking visual evidence can induce a self-evolving agent’s memory writer to create harmful source–action associations, causing persistent bad behavior even when the original evidence has been removed.

Critical strengths
  • Under a new memory attack surface. The paper shows how the visual evidence can induce the agent to write the harmful memory items, without injecting any target-specific content in attacker-controlled places.
  • Persistent effectiveness when the source has been removed. The attack remains effective after deleting the sources.
  • Detailed experiment demonstration. It covers two multimodal benchmarks, three memory mechanisms, and multiple models. Ablations, and defense evaluations further examine the attack mechanism and potential mitigations.
Critical weaknesses
  • It is unclear that how the target action will appear in normal use. In this paper, the attacker provides images without including the target action. But what makes the agent take the attacker desired action in the user tasks? It is not clear that how the tasks and feedback are used. How does the target action arise in normal use when the attacker only controls the source images? The commitment-flip test helps show when the association is formed, but it does not explain how the required action and feedback arise in ordinary tasks. It will be better if authors give a concrete example of the full interaction, showing what the attacker controls and why the agent takes the target action?
  • A zero success rate after cleanup does not mean the attack is gone. SC-ASR counts a success only when an attack memory is retrieved. Lineage Cleanup removes all such memories, so SC-ASR must be zero even if the agent still takes the target action. The paper should report Post-ASR after cleanup to check whether the behavior disappears. Without this, the reported zero does not establish that the derived memories are what sustain the attack.
Questions for authors
  • As I said in the weakness, what tasks and feedback lead the agent to take the target action? Please give one complete example, from the submitted image to the resulting memory, and explain what the attacker controls at each specific step.
  • Could you report Post-ASR after Source Delete, Lineage Cleanup, and Random Cleanup? This may show whether the target behavior disappears, rather than SC-ASR becoming zero because no attack-derived memory remains.
  • Could you test ELLA with later tasks sampled independently of the attack target? I want to see how often the target action occurs during memory evolution and the final Post-ASR.
4 comments
AuthorsOpenReview ↗

W1 / Q1. When the external provider controls only the image, how does the target action arise during ordinary interaction and enter memory?

We sincerely appreciate your thoughtful review and positive recognition of the memory-writing attack surface and the significance of persistent effects after source removal. Following your constructive suggestions, we have conducted additional analyses, including a complete interaction trace, direct post-cleanup behavioral measurements, and an evaluation on independently selected held-out tasks. We hope these results address your concerns and strengthen the evidence for ELLA.

We thank the reviewer for this insightful question and are pleased to provide a concrete, end-to-end interaction trace showing how the target action arises during an ordinary task and is subsequently consolidated into persistent memory, even when the attacker controls only the external image. Specifically, we examine an observed run with Qwen3-VL-30B-A3B + Mem0 + Mem-Gallery. An external provider submitted an ordinary photograph of a dog, a pet bed, and a water bowl. Neither the image nor its accompanying source context instructed the agent to select ALLOW or specified a source–action rule. A subsequent benchmark task asked the agent to select a safe action using its learned memory. The benchmark reference action was BLOCK. The agent selected ALLOW, which the evaluation environment recorded as executed, marked as incorrect, and assigned a reward of −1.0.

The Writer received the action and its execution status. The correctness judgment and negative reward were not passed to the Writer in this interaction. During consolidation, the Writer associated ALLOW with the dog-related source concept and stored a persistent memory entry linked to six external source records.

We believe this trace directly addresses the control-boundary concern. The external provider supplies only the image; the benchmark task provides the decision context; the agent selects the action; the evaluation environment judges the outcome; and the legitimate Writer consolidates source evidence and behavior into memory. Importantly, the attacker neither specifies ALLOW in the source nor directly writes the resulting association.

AuthorsOpenReview ↗

W2 / Q2. Does removing derived memories reduce how often the agent actually selects the target action?

We appreciate the reviewer's important distinction between lineage-conditioned success and actual agent behavior. In response, we additionally measured the conditional target-action rate after cleanup, without requiring retrieval of an attack-derived memory. All three cleanup conditions used the same 74 queries on which the paired Clean run did not select the target action. This metric counts action selections regardless of whether a derived memory was retrieved.

Table 1. Target-action rates under different cleanup conditions

Cleanup conditionVisual PFSVisual StaticPFS − Static
Source deletion28/74 (37.8%)18/74 (24.3%)+13.5 pp
Random cleanup27/74 (36.5%)17/74 (23.0%)+13.5 pp
Lineage cleanup9/74 (12.2%)8/74 (10.8%)+1.4 pp

We observe that lineage cleanup changes the agent's subsequent behavior, rather than merely forcing a lineage-based metric to zero. Compared with random cleanup, the target-action rate for Visual PFS drops from 36.5% to 12.2%, while the PFS–Static gap contracts from 13.5 pp to 1.4 pp. We acknowledge that some target-action selections remain, but the behavioral reduction directly supports the role of attack-derived memory.

We further examine whether the behavioral reduction merely reflects indiscriminate memory deletion. Under the same mean deletion budget of 2.6 records, lineage cleanup removes all designated derived records in this evaluation while preserving 100% of benign memories. We believe this selective removal provides complementary evidence that the intervention targets the relevant memory lineage.

Table 2. Memory retention and benign-task performance under a matched deletion budget

Cleanup conditionMean records deletedDesignated derived records remainingBenign-memory retentionBenign-task success
No cleanup0100%100%74%
Random cleanup2.6100%82.7%88%
Lineage cleanup2.60%100%80%
AuthorsOpenReview ↗

Q3. Do behavioral differences persist on subsequent tasks selected independently of the attack target?

We appreciate the reviewer's suggestion to evaluate tasks that are not designed around the attack target. We selected subsequent tasks from the existing Mem-Gallery held-out pool without using target-action labels or experimental outcomes. After fixing the task pool, we applied the paired Clean eligibility criterion, yielding 120 eligible queries per task category.

We first measure the target-action frequency during memory evolution: 11% for Visual PFS and 47% for paired Clean. These figures reflect immediate actions, whereas ELLA concerns how the Writer consolidates source–action associations for later reuse. We then measure subsequent behavior on the held-out tasks in Table 3.

Table 3. Conditional target-action rates on held-out tasks

Subsequent task categoryVisual PFSVisual StaticPFS − Static
Same domain, unrelated task objective28/120 (23.3%)13/120 (10.8%)+12.5 pp
Cross-domain, semantically related19/120 (15.8%)14/120 (11.7%)+4.2 pp
Cross-domain, semantically unrelated13/120 (10.8%)13/120 (10.8%)0.0 pp

We believe these findings show that ELLA's persistent influence is not restricted to later prompts explicitly requesting the attack target. Visual PFS exceeds Visual Static by 12.5 pp on same-domain tasks with unrelated objectives and by 4.2 pp on semantically related cross-domain tasks. No additional PFS–Static gap is observed in the semantically unrelated cross-domain group, suggesting that the influence is shaped by semantic relevance rather than unrestricted redirection. We will clarify the held-out sampling and subsequent Clean-based filtering in the revision.

We sincerely appreciate these constructive questions. We hope the additional interaction trace, behavioral cleanup results, and independent-task evaluation help resolve your concerns and encourage you to reconsider your assessment.

Area_Chair_PpxBOpenReview ↗

take part in discussion

Dear reviewer YbbB,

please take part in the discussion. time is running out.

thank you your ac

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

This paper introduces ELLA, a memory poisoning attack against self-evolving multimodal agents that exploits the memory formation process rather than directly injecting malicious instructions. The attacker provides benign-looking visual evidence that is likely to be retrieved during future interactions. When the agent subsequently performs an action, its memory writer may combine the retrieved evidence with the action and environment feedback, creating a persistent source–action association. The proposed method uses Semantic Anchoring and Writer Exposure to select effective visual evidence. Experiments across two multimodal benchmarks, three memory backends (Mem0, Graphiti, and A-MEM), and multiple reader models show that ELLA achieves higher attack success rates than existing memory poisoning baselines, with the effect persisting even after the original attacker-controlled content is removed.

Critical strengths

The paper presents an interesting perspective on memory poisoning by identifying the memory-writing process itself as an attack surface. Unlike conventional approaches that explicitly inject malicious instructions or associations, ELLA demonstrates how seemingly benign evidence can influence the formation of persistent memory. The evaluation covers multiple memory backends and reader models, suggesting that the vulnerability is not limited to a single implementation. The authors also conduct useful causal experiments, including source removal, lineage cleanup, and commitment-flip tests, to investigate whether the observed behavior originates from writer-generated memories rather than the original injected content.

Critical weaknesses

My main concern is whether the proposed attack reflects a realistic vulnerability in practical agent memory systems. ELLA assumes that the memory writer combines retrieved external evidence, agent actions, and environment feedback to form persistent source–action associations. However, many memory systems primarily extract facts, preferences, or conversation summaries rather than automatically associating retrieved content with subsequent actions. Although the paper evaluates Mem0, Graphiti, and A-MEM, it is unclear whether these associations emerge naturally from their standard memory pipelines or require a specifically configured memory-writing process. The No-Completion ablation also provides limited evidence for the proposed mechanism, since disabling memory formation would trivially prevent persistent poisoning.

Another major concern is whether the demonstrated attacks represent meaningful security threats rather than ordinary memory misalignment. The evaluated target actions, such as reserving a campsite or saving a recipe, are benign, and the paper defines attack success based on whether the agent performs an attacker-desired action rather than whether it violates user intent or causes harm. Since the target action originates from the agent's own ordinary decision-making process, the resulting association could simply reflect an incorrect generalization during memory consolidation. It would be more convincing to demonstrate scenarios where poisoned memory leads to unauthorized actions, violations of explicit user preferences, or other meaningful security failures. Without such evidence, the practical security implications of ELLA remain unclear.

Questions for authors
  1. Do the source–action associations emerge naturally from the default memory-writing pipelines of Mem0, Graphiti, and A-MEM, or do they require additional configuration?
  2. How do the authors distinguish malicious memory poisoning from ordinary memory misalignment or incorrect generalization?

I think the authors should explicitly have "ethics statement". The paper lacks a dedicated ethics statement and provides only a brief broader-impact discussion.

4 comments
AuthorsOpenReview ↗

W1 / Q1. Do practical memory systems form these source–action associations?

We sincerely appreciate your thoughtful review and positive comments on the memory-writing attack surface, cross-backend evaluation, and causal analyses, including source deletion, lineage cleanup, and Commitment Flip. We also appreciate your important questions about the realism of the memory-writing architecture and the security significance of the observed behavior. In response, we provide native writing checks and a new evaluation with explicit authorization constraints.

We appreciate the reviewer's concern that the observed associations might depend on a specially configured Writer. We believe the relevant threat model is not confined to an artificial application-specific rule, as our supplementary checks show that the native writing pipelines of Mem0, Graphiti, and A-MEM each formed source–action associations under the tested inputs. For the versions and inputs tested, no application-level rule constructed these associations in advance. In the main experiments, we held the application-level Writer fixed and varied the storage and retrieval backend. Table 4 presents the two sets of results.

Table 4. Native writing checks and cross-backend results with a fixed Writer

Memory systemQualitative native writing checkMem-Gallery SC-ASR (fixed Writer)MMRC SC-ASR (fixed Writer)
Mem0Source–action association observed53.7%47.2%
Graphiti (Kuzu in the cross-backend evaluation)Source–action association observed51.9%45.6%
A-MEMSource–action association observed62.0%60.8%

We further support this architectural premise with academic and industrial precedents. Reflexion distills reusable experience from feedback, and ExpeL does so from interaction trajectories. M3-Agent continually updates multimodal memories, while A-MEM creates or updates memory links as new experiences arrive [1–4]. MemEvolve and AdaMem further investigate memory architectures and writing policies [5, 6]. Amazon Bedrock AgentCore’s episodic memory organizes task goals, tool use, and execution outcomes. Zep Agent Skills uses interaction trajectories to produce reusable procedures. The public Iberdrola and OpenNote case studies also illustrate the practical need to reuse memory across interactions [7–10].

We then examine whether the subsequent association is actually generated during memory formation. In No-Completion, source evidence still reached the Writer. With consolidation disabled, no derived memories were retrieved and SC-ASR was zero. In Commitment Flip, we held the carrier image and pre-branch history fixed across 100 matched pairs and varied only the post-commit behavioral signal. Changing that signal also changed which action was more often activated later. We believe the combined evidence supports the central mechanism: action semantics can arise from the agent's subsequent interaction and be consolidated by the trusted Writer, even though the external image does not contain an explicit source–action rule.

Table 5. Paired results for Commitment Flip

Subsequently activated actionBehavioral signal a₁Behavioral signal a₂Paired difference (a₁ condition − a₂ condition)
a₁31.0%3.0%+28.0 pp
a₂3.5%32.5%−29.0 pp
AuthorsOpenReview ↗

W2 / Q2. How does ELLA differ from ordinary memory misalignment, and what authorization risk does it demonstrate?

We appreciate the reviewer's concern about distinguishing an adversarial memory-integrity attack from ordinary memory misalignment. We believe the key distinction lies in the attacker's deliberate influence across the boundary between lower-trust external evidence and trusted persistent memory. To further establish its security relevance, we conduct an additional evaluation with an explicit prohibition on selecting or executing ALLOW. Under this constraint, Visual PFS increased the confirmed unauthorized execution rate by 23.5 percentage points relative to Paired Clean (49.5% vs. 26.0%). Across 200 matched queries, we recorded target-action selections, requests for the prohibited action, and unauthorized executions confirmed by the execution logs.

Table 6. Requests and executions when ALLOW was explicitly prohibited

ConditionTarget-action selectionsUnauthorized requestsConfirmed unauthorized executions
Paired Clean52/200 (26.0%)52/200 (26.0%)52/200 (26.0%)
Visual Static88/200 (44.0%)88/200 (44.0%)88/200 (44.0%)
Visual PFS99/200 (49.5%)99/200 (49.5%)99/200 (49.5%)
PFS − Clean+23.5 pp+23.5 pp+23.5 pp
PFS − Static+5.5 pp+5.5 pp+5.5 pp

We believe these results strengthen the security interpretation of ELLA: an attacker who cannot edit the Writer or memory can still increase execution of a prohibited action in a controlled environment. Relative to Paired Clean, Visual PFS increases confirmed unauthorized execution by 23.5 pp. The observed PFS–Static difference is +5.5 pp. These execution findings complement source-deletion and lineage-cleanup experiments, which establish persistence and demonstrate that removing relevant derived memories reduces later target-action selections.

We further evaluate a separate execution-layer authorization check on 50 queries per group. It blocks all 24 unauthorized requests from Visual PFS and all 20 from Visual Static, yielding 0/50 confirmed unauthorized executions in both groups. We believe this result is particularly informative: persistent-memory manipulation can steer the agent toward prohibited requests, while an independent execution boundary can prevent these requests from becoming executed actions. We acknowledge that these are controlled authorization tests rather than evidence of compromise in a deployed system.

AuthorsOpenReview ↗

Q3. Dedicated Ethics Statement

We appreciate this valuable recommendation. We have included a dedicated Ethics Statement in the revised manuscript, reproduced below, describing our controlled evaluation, dual-use considerations, provenance tracking, execution-layer authorization checks, and auditable cleanup. We hope the additional native-writing and authorization experiments, together with these clarifications, address your concerns about realism and security impact, and encourage you to reconsider your evaluation.

Ethics Statement in the Revised Manuscript

This work examines how multimodal agents with long-term memory integrate external evidence into persistent state and how that state affects subsequent decisions. All experiments were conducted in controlled, isolated evaluation environments. In the authorization experiments, violations were identified using predefined task rules and execution logs. The results apply to the reported evaluation settings. Manipulating persistent memory has dual-use implications. We recommend preserving memory provenance and feedback from the environment, checking authorization at execution time, and providing auditable mechanisms for removing derived memories. Future evaluations of safeguards should measure both security outcomes and performance on benign tasks.

References

[1] Shinn et al. Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS 2023.

[2] Zhao et al. ExpeL: LLM Agents Are Experiential Learners. AAAI 2024.

[3] Long et al. Seeing, Listening, Remembering, and Reasoning: A Multimodal Agent with Long-Term Memory. ICLR 2026.

[4] A-MEM: Agentic Memory for LLM Agents. 2025.

[5] MemEvolve: Meta-Evolution of Agent Memory Systems. ICML 2026.

[6] Chen et al. AdaMem: Learning What to Remember with Adaptive Memory Policies for Personalized Agents. EMNLP Industry Track 2026.

[7] AWS. Amazon Bedrock AgentCore: Episodic Memory Strategy. Official documentation.

[8] Zep. Agent Skills: Capture Traces. Official documentation.

[9] AWS. Iberdrola Enhances IT Operations Using Amazon Bedrock AgentCore. Customer case study.

[10] Mem0. How OpenNote Scaled Personalized Visual Learning with Mem0. Customer case study.

Area_Chair_PpxBOpenReview ↗

take part in discussion

Dear reviewer HthB,

please take part in the discussion! time is running out.

thank you very much your ac

Comments on the paper

2 comments

Dear Reviewers,

We sincerely thank all reviewers for their time and constructive comments. We greatly appreciate the opportunity to further strengthen our work through this discussion.

We have provided detailed point-by-point responses to each reviewer and conducted additional experiments and analyses to address the reviewers' concerns. Our additional findings show that source–action associations can emerge through native memory-writing pipelines, persist beyond the removal of attacker-controlled sources, and influence subsequent agent behavior. Further cleanup and authorization experiments provide evidence of their causal role and security implications. Beyond these empirical results, we have further elaborated on the key insight behind ELLA's design: selecting ordinary visual evidence that reaches the memory writer can induce persistent associations without explicitly injecting the target action. These findings further support our methodological choices and the significance of the identified security risk.

We hope our detailed responses and additional findings adequately address the reviewers' concerns and further substantiate the significance of our work. We sincerely value this opportunity to address your questions and engage in constructive discussion. We look forward to your further feedback and are eager to provide additional explanations or evidence that may help resolve any remaining concerns.

Thank you again for your time and consideration.

AuthorsOpenReview ↗

General Response

As the discussion will soon come to a close, we would like to further clarify our work and reiterate its central message. We sincerely appreciate the reviewers’ thoughtful questions and constructive suggestions. We are also grateful for their recognition of several strengths of ELLA:

  • The importance of memory formation as an underexplored attack surface in self-evolving agents.
  • The mechanism through which source-only visual evidence can induce writer-generated associations without specifying the target action.
  • The persistence of the induced effects and the value of our causal analyses.
  • The breadth of our evaluation across memory backends and reader models.

1. Further Clarification on the Main Concerns

The main concerns have focused on the realism of memory formation, the causal origin and persistence of the induced associations, and their security implications. We further examined these aspects:

  • Realism: Separate qualitative checks of the native writing pipelines observed source–action associations in Mem0, Graphiti, and A-MEM without application-level rules preconstructing them in the tested settings. An end-to-end trace illustrates how source-only visual evidence, the agent’s own action, and the writer’s update come together in an evaluated task.
  • Causality and persistence: In No-Completion, source evidence reached the writer in all measured interactions (100% writer exposure), yet PRR and SC-ASR fell to zero when consolidation was disabled. Commitment Flip shows that the activated action follows the post-commit behavioral signal while the carrier and pre-commit history remain fixed. Source/lineage cleanup and held-out-task evaluations further examine persistence and subsequent behavioral effects.
  • Security implications: Under an explicit prohibition on ALLOW, Visual PFS produced 49.5% confirmed unauthorized executions, compared with 26.0% for Paired Clean (+23.5 percentage points). We also evaluated execution-time authorization and writer-side defenses, and added a dedicated Ethics Statement.

2. The Perspective and Significance of ELLA

We believe ELLA is a solid piece of work because it investigates an important security question in self-evolving agents: Can an attacker influence an agent’s future behavior by supplying only source-side visual evidence, while the final association is formed during the agent’s own memory update?

Our work focuses on how external source evidence and the agent’s subsequent behavior become linked during memory formation. The attacker supplies source-side visual evidence; the target action emerges from the agent’s subsequent task interaction; and the trusted writer brings these signals together into a persistent association. We refer to this boundary as the semantic-completion point. Our central insight is that the victim’s own memory writer can complete an association that is not explicitly specified in the attacker-provided evidence.

This insight shapes ELLA’s methodology. Semantic Anchoring (SA) places source evidence in the semantic neighborhood of future source queries, while Writer Exposure (WE) uses bounded public feedback to increase the likelihood that this evidence reaches the writer. These mechanisms establish the conditions for reach; the association is formed through the victim’s subsequent behavior and memory update. ELLA thus connects limited attacker control over external evidence with the victim-side process that creates persistent memory.

We provide a systematic investigation of where the association is completed, how source-only evidence reaches the writer, and whether the resulting memory persists and influences subsequent behavior.

Our evaluation shows that ELLA achieves the highest SC-ASR in all six matched benchmark–backend settings, exceeding the strongest baseline by 21.6–28.0 percentage points. Cross-reader evaluations, source-removal and lineage interventions, and authorization experiments further characterize the transferability, persistence, and security implications of the induced associations. After attacker-controlled sources are removed, macro-average PRR and SC-ASR remain at 81.5% and 55.9%, respectively.

We believe ELLA brings memory formation into the security analysis of self-evolving agents. As these agents increasingly accumulate and reuse experience, their security also depends on how memory writers interpret, associate, and preserve external evidence across interactions.

We sincerely thank the reviewers and Area Chairs for their time, feedback, and recognition of our work’s strengths. We hope this clarification conveys ELLA’s research question, underlying insight, and systematic investigation, and that these aspects will be considered in the ongoing assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.