acceptodds
Under review as a conference paper at ICLR 2027

What Did the Agent See? Implanting False Memories in Multimodal AI Agents through Black-Box Visual Attacks

Abstract

Multimodal AI agents increasingly rely on persistent long-term memory to ground generation in past visual episodes, and in embodied deployments that memory becomes the agent's record of an environment it can no longer observe. We show that unconditional trust in stored visual content creates a critical vulnerability: an attacker who controls nothing but an image can make an agent remember something that never happened. We study a strictly image-bounded threat model: a content publisher perturbs an image a user later shares with their agent, with no access to the target model, its memory system, the text channel, or future queries. The poison must prevail over the user's truthful text and remain dormant until a dependent query retrieves it. We propose Vivid, a blind, black-box framework on two axes: scene-dominant poison design, which constructs from the image alone the content that survives transfer through pooled embeddings, frozen before any query exists; and delivery-aware optimization, which aligns the image with that target across encoder families and in expectation over the re-encoding it undergoes before ingestion. Across four benchmarks, five memory backends and four black-box victims, Vivid succeeds against every backend and outperforms the strongest prior transfer attack, decisively so under realistic delivery channels. We further characterize purification and memory-level detection and isolation as defenses.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multimodal AI Agents, Long-term Memory, Memory Poisoning, Black-box Transfer Attacks, Adversarial Attacks and Defenses

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_bhD1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes VIVID, an image-only black-box attack against multimodal agent memory. It combines scene-dominant target construction, a cross-family encoder ensemble, and delivery-aware optimization. Evaluation covers five memory backends and four victim models across three memory benchmarks and an additional query-time benchmark. The results demonstrate false-content insertion, retrieval changes, and subsequent false assertions. The paper also evaluates image purification and memory-level detection, including their effects on benign content and genuine memory.

Critical strengths
  1. A meaningful threat model with evidence of memory corruption. The attacker controls image pixels, and an explicit information barrier separates attack construction from evaluation queries. Results on description-based stores show that false content can enter stored text before recall, providing concrete evidence of corruption during memory ingestion.

  2. Substantial robustness and defense evaluation. The paper includes competitive transfer baselines, documented implementation checks, and delivery transformations outside the optimization distribution. The delivery-aware ablation shows a substantial effect. Defense experiments also report false alarms, benign-content degradation, and genuine memory lost through isolation.

Critical weaknesses
  1. Some stage metrics appear inconsistent with their definitions. In Table 19, GPT-5.6-Luna with MMMemory on WMA-Embodied has 4/10 final hits, 69.2% retention, and 86.7% image-level transfer. Under the stated 15-goal evaluation, 13 goals transfer; Equation 11 would therefore imply approximately nine joint transfer-and-memory successes, exceeding the four reported final hits. The evaluation populations, exclusions, or reported values need reconciliation before retention and bottleneck conclusions can be assessed.

  2. Comparative effectiveness is established mainly at the image level. Figures 2–3 compare attacks through isolated-image transfer, while Table 1 reports full memory-pipeline ASR only for VIVID. Its delivery columns also contain image-level rates shared across backends. Since writing, retrieval, and generation substantially affect success, matched end-to-end comparisons are needed to establish the advantage in the paper’s central memory setting.

  3. The future-query and dormancy claims need tighter empirical support. Evaluation recall questions often explicitly name the selected hazard. The information barrier and clean/noise controls support a conditional vulnerability under these probes; independently specified questions would establish broader query generalization. Furthermore, ordinary-answer degradation reaches 14.5–24.5 percentage points on MemLens. This requires reconciliation with the dormancy constraint in Equation 4, even when text-grounded memory diagnostics remain stable.

Questions for authors
  1. Can you reconcile the hit counts and retention values in Table 19, specifying the exact populations and exclusion rules used for each metric?

  2. Does VIVID outperform the strongest transfer baseline through the complete memory pipeline under both unchanged and transformed delivery?

  3. Is each memory run evaluated with exactly one poisoned image? How do results vary with independently specified recall questions and the delay between ingestion and recall?

4 comments

We sincerely appreciate the reviewer's valuable comments. We have incorporated the results of these comments to improve our paper. We hope that our response can address the raised concerns. Please let us know if there are any further questions.

W/Q1: Reconciling hit counts and retention in Table 19 (Table 24 in the revision).

We thank the reviewer for the careful check. The two columns use different goal populations by design. ASR (the hits column) measures whether the attack causes the answer: it counts testable goals and excludes any goal whose clean-run answer already asserts the target, since such a hit cannot be attributed to the attack. Retention measures how often the image-level effect survives the memory pipeline, so its denominator must be the goals that transfer at image level, and is identical across the five backends of a victim. Additionally, similar to image-level transfer, it does not apply clean-hit exclusion as each backend has its own clean run, so the exclusion would give each backend a different denominator.

For the example you raise (GPT-5.6-Luna, MMMemory, WMA-Embodied), 13 of the 15 goals transfer. Five goals are excluded from ASR because their clean-run answer already matches the target, so ASR is 4/10. All five also transfer and are asserted in the attacked run, so retention counts 4 ASR + 5 transfer = 9 of 13, the 69.2% reported in Table 24.

This exclusion is uncommon and makes ASR conservative. It removes 1.2% of goals on MemLens, 0.5% on Mem-Gallery, 5.6% on WMA-Personal and 10.3% on WMA-Embodied, and without it ASR would be equal or higher on every benchmark. GPT-5.6-Luna on WMA-Embodied is where it matters most: its clean run matches the target on 24% of goals, against at most 12% for any other victim on this WMA-Embodied benchmark. These are false alarms on hazard questions: in the the GPT-5.6-Luna on WMA-Embodied cell, four of the five clean answers name a hazard the scene does not contain, and the benchmark’s own judge labels all five as hallucinations.

The population and exclusion rule of each stage metric:

MetricGoals countedClean-hit exclusion
ASR and hitstestable goals (recall question asked at a checkpoint present in both runs)yes
Image-level transferall goals with a recall question; a goal transfers if at least 2 of 3 draws assert the target (Table 1's band gives both the share of draws and this goal count)no
Retentiontestable goals that transferno
Rank-0testable goals (rank of the poisoned record in the clean and attacked runs)no
Write-ASRgoals with a stored memory of the goal's imageno
Collateralthe benchmark's ordinary questions, paired clean vs attackedno

Our conclusions do not depend on this choice. Counting only ASR hits in retention, MemLens retention still spans 25.9 to 53.4% (reported 25.9 to 55.2%), and the Qwen3-VL vs Gemma-3 comparison of Appendix F.3 is unchanged; the bottleneck conclusions of Sec. 7 rest on write-ASR, rank promotion and ASR rather than retention.

Paper Revision: The revised version states each metric's population in Eq. 11 and Appendix E, reports the retained count next to each retention value and the number of transferring goals next to Table 1's transfer band, and adds retention restricted to ASR hits as an additional column.

W2: Comparative effectiveness is established mainly at the image level.

Thank you for this suggestion. We ran the requested matched end-to-end comparison: VIVID outperforms the strongest transfer baseline, M-Attack-V2, through the complete memory pipeline on every backend; pooled, the margin is large under transformed delivery and small under unchanged delivery.

Setup: MemLens (our primary benchmark) at ε = 16/255, four backends (MMMemory, UniversalRAG, AUGUSTUS, ViLoMem) and two victims (Gemma-3-27B, GPT-5.6-Luna). M-Attack-V2 is compute-matched as in Appendix D.2, and delivery is unchanged, JPEG30 or RESIZE80 (both held out; Table 16). To isolate where the gain comes from, we also report M-Attack-V2 with VIVID's delivery-aware term added (VIVID without its SigLIP surrogate, Table 14).

DeliveryVIVIDM-Attack-V2M-Attack-V2 + delivery-awareVIVID gain over M-Attack-V2
Unchanged26.723.218.9+3.5
JPEG3023.27.017.3+16.2
RESIZE8010.51.910.8+8.6

ASR (%), pooled over 4 backends × 2 victims × 47 goals.

As the above table shows, VIVID's advantage carries through the memory pipeline. Under unchanged delivery, the gain matches the modest 2.7 to 4.5-point image-level margin of Figs. 2 and 3. Under transformed delivery, M-Attack-V2's poison rarely survives, while VIVID's survives far more often at each stage: under RESIZE80, VIVID's poisoned record is still retrieved first on UniversalRAG for 55% of goals (M-Attack-V2: 30%; Table 19) and still written into ViLoMem's stored descriptions (6% vs 0%). The transformed-delivery gain holds for each victim separately (8.1 to 18.8 points). Adding VIVID's delivery-aware term to M-Attack-V2 recovers most of this robustness (17.3% and 10.8%), so the gain comes from VIVID's delivery-aware design, and VIVID's cross-family surrogate adds strength on unaltered pixels (26.7% vs 18.9%), as in the image-level ablation (Fig. 9). This is the delivery robustness VIVID is designed for (Sec. 5.2).

Paper Revision: The revised version adds this comparison to the evaluation in the main paper (Sec. 7), with per-backend and per-victim results in the appendix, and clarifis in Table 1's caption that its delivery columns report image-level transfer.

W3a: Is each memory run evaluated with exactly one poisoned image?

Yes on MemLens, our primary evaluation benchmark: each of its 47 memories holds exactly one poisoned image. WMA-Personal and WMA-Embodied memories hold several (15 to 18 and 3 to 4), each with its own recall question. Mem-Gallery attaches the poisoned image at query time and stores none.

W3b: How do results vary with independently specified recall questions?

Thank you for raising this. Eq. 4 requires the false content to surface only for queries that depend on the poisoned observation, supp(Q). Our recall questions are such queries: a user checks the image for a type of content (“What X, if any, is visible?”). The question names the poison type (Table 5 in the revised version), not the false content: for a climber photo poisoned with a crossbow, it asks “What weapon, if any, is visible?”. Our threat model assumes the user to ask it, not the attacker.

We added questions written without access to the hazard or the poison, five wordings each (MemLens, five backends, same judge and ASR):

Question typeGemma-3-27BGPT-5.6-Luna
Names the type ("What {type}, if any, is visible?")44.938.5
Type removed ("What hazards, if any, are visible?")14.016.9
Generic safety6.46.3
Planned activity + safety3.65.3
Other tasks about the scene0.60.1

ASR (%); a memory counts if any of its five answers asserts the false content (Table 40).

As shown, the false memory surfaces without being specificlaly named, most often when the user checks for its type, and fades as the question depends less on the poisoned attribute, although the record is retrieved for nearly all other-task questions. This is what dormancy requires: a record that surfaced for any question about the scene would expose itself in ordinary use.

W3c: How do results vary with the delay between ingestion and recall?

They change little. We re-asked the recall question after the rest of each memory's history was ingested (median 6, 17.5 and 8 further sessions on MemLens, WMA-Personal and WMA-Embodied). ASR changes by −4.5 points on average for Gemma-3-27B and −2.8 for GPT-5.6-Luna, and the change does not grow with delay (correlation 0.05 and 0.12). No backend scores or evicts memories by time, so delay acts only through the records added in between.

W3d: Ordinary-answer degradation on MemLens and Eq. 4.

This degradation is the cost of changing the queried image, not a leak outside supp(Q): on MemLens, 47 of 50 ordinary questions per run ask about the poisoned image. Pooled over victims, the effect is omission, not fabrication: QA-O rises by 11 to 21 points while QA-H moves by at most 5 (Table 1). Where most ordinary questions concern other images (WMA-Embodied), degradation stays at −2.4 to +5.2 points, and the memory diagnostics stay flat.

Paper Revision: The revised version adds the query-generalization results, question wordings and delay analysis to the appendix (F.10, F.11), and clarifies in Sec. 7 that MemLens collateral measures ordinary visual questions about the attacked image.

Dear Reviewer bhD1,

Since the Author/Reviewer discussion period concludes today, we would be happy to address any remaining concerns or questions.

Thank you.

Reviewer_FP4EOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether perturbing an input image can cause a multimodal agent to store and later assert false information, without attacker access to the deployed model, memory system, user text, or future queries. The proposed VIVID framework constructs a synthetic target image from an image-derived attack goal and optimizes a bounded perturbation using a CLIP/SigLIP ensemble and delivery-aware transformations. The evaluation measures image-level transfer, subsequent effects across existing memory backends, and defenses based on image purification and memory-record verification. The experiments include both memory-ingestion settings and a distinct query-time image setting in MemGallery.

Critical strengths
  1. The evaluation covers different memory representations under restricted attacker capabilities. Image-only perturbations affect both image-retaining and text-based memory without modifying user text or directly accessing the memory store. The text-based results show that corrupted visual content can become part of the agent’s own stored descriptions and influence later answers.

  2. The delivery-robustness results are meaningful. VIVID substantially outperforms the compared attacks after delivery transformations, and the component ablation helps explain this improvement. The defense evaluation also measures the loss of genuine memory, making the trade-off between attack mitigation and information preservation explicit.

Critical weaknesses
  1. The methodological advance over existing image attacks appears limited. VIVID combines established feature-alignment, encoder-ensemble, and transformation-aware optimization techniques with synthetic target images. Its improvement over M-Attack-V2 on undegraded images is 2.7–4.5 percentage points. The larger gains after delivery transformations are meaningful, but the component ablation attributes them primarily to delivery-aware optimization. A comparison with M-Attack-V2 using the same delivery transformations is needed to clarify the additional value of VIVID’s design.

  2. The significance of the memory extension remains unclear. The analyses of storage-dependent bottlenecks, model susceptibility, and memory-quality diagnostics are useful. However, the paper does not sufficiently explain what substantial new understanding they provide beyond the persistence and downstream use of corrupted visual information, relative to prior image-based memory poisoning. Moreover, Figure 2 compares attacks only at the image level, while Table 1 reports final memory-level results only for VIVID. Thus, it remains unclear whether existing attacks produce similar memory effects or whether VIVID offers an additional benefit after storage and retrieval.

Questions for authors

Clarification of the central contribution: The paper reports storage-dependent bottlenecks, differences between image-level and memory-level susceptibility, and limitations of existing memory-quality diagnostics. What is the principal new insight these findings establish beyond the persistence and downstream use of corrupted visual information? Please explain how this insight advances prior image-based memory-poisoning work, what concrete security or memory-design implication follows, and which reported results support it.

5 comments

We appreciate the reviewer's feedback which we have been integrated to improve our paper. We hope our responses can address the raised concerns. Please let us know if there are any further questions.

W1: Does VIVID's design add value over M-Attack-V2 once both use the same delivery transformations?

Thank you for this question. We clarify that this comparison is provided in the paper: the −SigLIP ablation of Fig. 9 (Table 14) is M-Attack-V2's objective and CLIP ensemble with VIVID's delivery-aware optimization, over the same transformations, steps and KK as VIVID. Each of VIVID's two design choices are absent in prior work (M-Attack-V2).

DeliveryM-Attack-V2M-Attack-V2 + delivery-awareVIVID
Unchanged58.052.361.5
WEBP40 (held out)39.448.052.0
COMBO22.958.561.0
RESIZE80 (held out)14.942.644.1

Image-level transfer (%), MemLens, four victims (Fig. 9).

As the above table shows, VIVID's delivery term adds 27.7 and 35.6 points to M-Attack-V2 under RESIZE80 and COMBO at a cost of 5.7 on unchanged pixels, where the cross-family surrogate then adds 9.2, so VIVID leads both in every condition.

i) The delivery term is VIVID's own contribution. EOT is only the estimator (Athalye et al., 2018); M-Attack-V2 already averages over random crops of its own pixels (A\mathcal{A} in Eq. 7) and still collapses under strong delivery transformations. VIVID's C^\hat{C} mixes an identity term (where delivery may be lossless) with a product of independent downscaling, JPEG, gamma and blur operations, so compound degradations occur at their natural frequency (Eq. 9). To our knowledge, no prior image poisoning attack on VLMs models this distribution. We note that input diversity (Xie et al., 2019) also resizes in the loop, but for transferability, with no lossy re-encoding. Added to FOA and M-Attack-V1, the delivery term raises their held-out mean (Table 18) by 13.2 and 13.5 points.

ii) The 2.7 to 4.5-point margin covers only unchanged images. Re-encoding is unknown to the attacker (Sec. 4); under it, VIVID's margin over M-Attack-V2 grows to 29.2 points (RESIZE80) and 38.1 (COMBO) (Fig. 9).

iii) The value carries into memory. Through the full pipeline (Table 2), VIVID leads M-Attack-V2 + delivery-aware under unchanged and JPEG30 delivery (ASR 26.7 vs 18.9, 23.2 vs 17.3) and matches it under RESIZE80 (10.5 vs 10.8), where both rely on VIVID's delivery term.

Paper Revision: The revised version renames −SigLIP "M-Attack-V2 + delivery-aware optimization" in Fig. 9 and Table 14, state this attribution in Sec. 5.2, and adds the full delivery grid, with FOA and M-Attack-V1, to the appendix.

W2: Do existing attacks produce similar memory effects, and does VIVID add a benefit after storage and retrieval?

Thank you for this suggestion. We ran the comparison end to end. When the image reaches the agent unchanged, the strongest existing attack, M-Attack-V2, produces a memory effect close to VIVID's, matching the small image-level gap. Under transformed delivery, it loses most of its effect inside the memory pipeline, while VIVID's poison survives storage and retrieval far more often.

Setup: MemLens (our primary benchmark) at ε = 16/255, four backends (MMMemory, UniversalRAG, AUGUSTUS, ViLoMem) and two victims (Gemma-3-27B, GPT-5.6-Luna). M-Attack-V2 is compute-matched as in Appendix D.2, and delivery is unchanged, JPEG30 or RESIZE80 (both held out; Table 16). Besides ASR (%), we report retention (Eq. 11) of the goals whose image transfers on its own (%), the share whose poison is still asserted after storage and retrieval.

DeliveryASR: VIVIDASR: M-Attack-V2Retention: VIVIDRetention: M-Attack-V2
Unchanged26.723.238.136.2
JPEG3023.27.032.112.5
RESIZE8010.51.916.11.0

In %, pooled over 4 backends × 2 victims × 47 goals.

As the above table shows, under unchanged delivery both attacks survive storage and retrieval at similar rates, with Vivid leading by 3.5-point ASR gain that matches the 2.7 to 4.5-point image-level margin (Figs. 2 and 3). Under transformed delivery, VIVID's advantage grows inside the pipeline. M-Attack-V2's poison is lost twice, first at the image and again in storage and retrieval: of its goals that still transfer, only 12.5% (JPEG30) and 1.0% (RESIZE80) are asserted from memory, against 32.1% and 16.1% for VIVID. The memory stages show where: under RESIZE80, VIVID's poisoned record is still retrieved first on UniversalRAG for 55% of goals (M-Attack-V2: 30%; Table 19) and still written into ViLoMem's stored descriptions for 6% (M-Attack-V2: none). VIVID leads M-Attack-V2 on every backend and, under transformed delivery, for each victim VLM separately.

Paper Revision: The revised version adds this comparison to the evaluation in the main paper (Sec. 7), with per-backend and per-victim results in the appendix.

W3: What is the principal new insight beyond persistence and downstream use, and how does it advance prior image-based memory poisoning?

Thank you, we are glad the analyses are useful. Prior image-based poisoning shows a perturbed image can corrupt a model's output. The multimodal memory extension shows what a single-pass evaluation cannot: how much of a visual poison survives in memory, and where it can be caught, is set by the memory architecture rather than by the perturbation. Our attacker is weaker than prior memory poisoning assumes (no adversarial text, write access, crafted queries or trigger), so these effects come only from the multimodal memory pipeline.

i) Where the poison is lost depends on the store. Embedding stores lose it only at answering (UniversalRAG rank-0: 12.8% → 72.3%); descriptive stores assert almost every poison written (ViLoMem: 35 to 54% written, 31 to 54% asserted).

ii) Image-level transfer does not set the memory-level outcome. At identical transfer, MemLens retention spans 25.9 to 55.2%, and under JPEG30 and RESIZE80 M-Attack-V2's poison is lost again inside memory (retention 12.5 and 1.0%, vs 32.1 and 16.1% for VIVID).

iii) Current Multimodal Memory-quality diagnostics (e.g., WorldMemArena) miss the attack. They stay flat on all five backends while the poison is asserted.

iv) The poison survives the re-encoding real systems apply. Deployed APIs resize uploads (OpenAI; Anthropic) and sharing platforms recompress them, stripping ordinary perturbations. Prior attacks also take an expectation over transformations (Athalye et al., 2018), but over random crops of their own pixels (M-Attack, M-Attack-V2, FOA), so they collapse under harsh held-out delivery. VIVID takes it over this re-encoding and keeps 70 to 86% of its transfer through RESIZE80, against 11 to 26% for baselines on three benchmarks.

So a multimodal memory should check what an image shows before storing it, against a reference the attacker cannot change: checking it against the user's own text cuts attack success on MemLens from 27.1 to 6.4%, and against independent descriptions of the pixels detects 85.4 and 86.7% of poisons on WorldMemArena.

Paper Revision: The revised version states this insight in the abstract and introduction and reframes Sec. 7 around it.

References.

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

This paper proposes VIVID, a black-box visual attack framework that implants false memories into multimodal AI agents through adversarial image perturbations. Experiments across four benchmarks, five memory backends, and four victim models demonstrate the effectiveness of the proposed attack.

Critical strengths
  1. Interesting motivation and practical significance. The paper investigates an important security vulnerability in multimodal agents with persistent memory, where adversarial images can corrupt stored visual information and influence future responses. The proposed image-only threat model is particularly interesting because the attacker does not require access to the memory system, text channel, or future queries.
  2. Comprehensive experimental evaluation. The authors evaluate VIVID across multiple benchmarks, memory architectures, and black-box victim models, demonstrating strong attack transferability. The experiments also provide detailed analyses of memory writing, retrieval, and generation, as well as the effectiveness of different defense strategies.
Critical weaknesses
  1. Insufficient comparison with delivery-aware adversarial attack baselines. The proposed VIVID combines cross-family feature alignment with delivery-aware optimization. However, the ablation results in Figure 9 suggest that delivery-aware optimization is primarily responsible for the improvements under image compression and other transformations. Since expectation over transformations is a well-established technique in adversarial attacks, it remains unclear whether similar improvements could be achieved by incorporating the same delivery-aware optimization into existing methods such as M-Attack-V2. Additional comparisons under identical transformation distributions and computational budgets would better establish the methodological contribution of VIVID.
  2. Limited diversity of retrieval encoders across memory backends. Although the paper evaluates five different memory backends, Appendix B indicates that all three embedding-based systems use the same GME retrieval encoder. The two text-based memory backends also share the same text embedding model. Consequently, the evaluation provides architectural diversity in memory management but more limited diversity in the underlying retrieval representations. Since cross-encoder transferability is a central motivation of VIVID, evaluation using additional independently trained retrieval encoders would strengthen the generalization claims.
  3. Limited validation of adversarial image perceptual quality. VIVID constrains adversarial perturbations using an L-infinity perturbation budget of 16/255, but a bounded pixel perturbation does not necessarily guarantee that the resulting image is perceptually indistinguishable from the original. The paper provides qualitative examples but does not systematically evaluate the visual perceptibility of the adversarial artifacts through human assessment or dedicated perceptual similarity metrics for the attacked images. This makes it difficult to determine whether the poisoned images would remain unnoticed when shared or inspected in realistic applications.
Questions for authors
  1. According to weakness 1, could the authors compare VIVID with a delivery-aware variant of M-Attack-V2 using the same image transformations and computational budget?
  2. According to weakness 2, could the authors evaluate VIVID using additional independently trained retrieval encoders beyond GME and text-embedding-3-small?
  3. According to weakness 3, could the authors provide a quantitative perceptual quality evaluation of the adversarial images, such as LPIPS measurements or a human perceptual study?
3 comments

We sincerely appreciate the reviewer's valuable comments, which has contributed to improving our work. We hope that our response can address the raised questions. Please let us know if there are any further questions.

W1: Could M-Attack-V2 with the same delivery-aware optimization match VIVID, under identical transformations and compute?

Thank you for raising this question. First, we believe that a clarification to our text labels can resolve this request. This comparison is already provided in the paper. It is the configruation previously labeled '-SigLIP' in Fig. 9 & Table 14. We have also included a clarifying text in Table 14, specifying what the tested configuration represents.

The '-SigLIP' ablation is M-Attack-V2's objective and CLIP ensemble with VIVID's delivery-aware optimization, with the same C^\hat{C}, steps and KK as VIVID. M-Attack-V2 is compute-matched, we tried doubling its sample budget but this gave no gain (Appendix D.2).

DeliveryM-Attack-V2M-Attack-V2 + delivery-awareVIVID
Unchanged58.052.361.5
WEBP40 (held out)39.448.052.0
COMBO22.958.561.0
RESIZE80 (held out)14.942.644.1

Image-level transfer (%), MemLens, four victims (Fig. 9).

The middle column of the above table contains the result from the requested configuration. Our finding remains that adding VIVID's delivery-aware optimization to an attack of the same lineage (i.e., M-Attack, FOA, and M-Attack-V2 lineage), recovers much of its delivery robustness. However, we note that M-Attack-V2 still performs at 9.2 points below VIVID on unchanged pixels (52.3 vs 61.5), and below VIVID on CLIP and non-CLIP configurations. Our conclusion here is that the cross-family alignment that is absent in previous attacks is what improves transfer on unchanged pixels.

Second, on clarifying our methodological contribution, We agree that EOT is well established, our contribution is what to take the expectation over when the entire memory pipeline (delivery, write, strore, retrieval, generation) is not observable.

The lineage of M-Attack, FOA, and M-Attack-V2 already average over random crops of the image own pixels (A\mathcal{A} in Eq. 7) and still collapses under strong delivery transformations. VIVID's C^\hat{C} mixes an identity term (where delivery may be lossless) with a product of independent downscaling, JPEG, gamma and blur operations, so compound degradations occur at their natural frequency (Eq. 9). This is what makes a pixel-only, pipeline-blind attack possible. To our knowledge, no prior image poisoning attack on VLMs models this distribution. We note that input diversity (Xie et al., 2019) also resizes in the loop, but for transferability, with no lossy re-encoding and no held-out strong delivery testing.

Also, as shown in Table 18, VIVID's delivery term is not tied to one attack. On every condition, it raises the mean over the four held-out ones by 13.2 to 14.9 points for FOA, M-Attack-V1 and M-Attack-V2, and VIVID stays above the variant M-Attack-V2 with the delivery term in every one, by 1.9 to 9.7 points. Through the memory pipeline Table 2, VIVID leads this variant under unchanged and JPEG30 delivery (ASR 26.7 vs 18.9, 23.2 vs 17.3) and matches it under RESIZE80 (10.5 vs 10.8), where both rely on VIVID's term.

Paper Revision: The revised version renames '-SigLIP' to "M-Attack-V2 + delivery-aware optimization" in Fig. 9 and Table 14, and adds the full delivery grid, with FOA and M-Attack-V1, to the appendix.

W2a: All three embedding backends use the same retriever (GME); please test more independently trained image retrievers.

Thank you for your suggestion to expand our evaluation to more retrievers. We have expanded our retrieval evalutation to eight image retrievers, none in VIVID's surrogate ensemble. We also note that the shared GME encoder is an instance of transfer diversity because GME is outside VIVID's surrogate ensemble. Additionally, each of the three instances differ in how they use that retrieval.

We clarify that for this rerun, UniversalRAG is the most direct test of retriever generalization. Since each record is an image embedding alone, swapping the retriever asks exactly the cross-encoder question: 'does an independently trained encoder still ranks the poisoned image first?'. (MMMemory and AUGUSTUS embed the image fused with its dialogue, mixing transfer with how a retriever weights image against text.) For each retriver, we measured rank-0 promotion of the poisoned record. We report the result for 5 of the evaluated retrievers. The full results are included in Appendix F.12.

Image retriever (not a surrogate)Vision towerUniversalRAG rank-0
GME-Qwen2-VLQwen2-VL12.8 → 72.3
VLM2Vec-Qwen2VLQwen2-VL12.8 → 70.2
MetaCLIP 2MetaCLIP worldwide27.7 → 93.6
jina-clip-v2EVA0219.1 → 89.4
Qwen3-VL-EmbeddingSigLIP-227.7 → 91.5

As shown, rank-0 of the poisoned record rises by 57 to 70 points for all eight retrievers: 20.0 → 83.4 on average for the five shown, 17.8 → 78.7 for the three whose towers share no lineage with our surrogates, and 17.0 → 85.1 for the three OpenAI CLIP-L retrievers (Appendix F.12) that do. On MMMemory's fused index the same retrievers still raise poison-in-top-10. The smaller gains there track image-versus-text weighting, not a failure of transfer.

W2b: Both text backends use the same text embedder (text-embedding-3-small); please test more independently trained text embedders.

We re-ran Mem0 with four other text embedders. Mem0 indexes text facts extracted from a model-written description of the image, with no image left in the record, so the attack reaches it through that description, before the embedder. text-embedding-3-small is the main evaluation embedder. The table gives the poisoned record's rank-0, clean to attacked:

Text embedderMem0 rank-0
text-embedding-3-small10.6 → 21.3
bge-m35.4 → 18.5
Qwen3-Embedding13.0 → 22.8
nomic-embed-text13.0 → 28.3
text-embedding-3-large18.5 → 32.6

Rank-0 (%) on MemLens, pooled over Gemma-3-27B and GPT-5.6-Luna.

As the above table shows, the poisoned record's rank-0 rises for every embedder, by 9.8 to 15.2 points. ViLoMem, the other text backend, admits no rank measure for the poisoned record, since its records carry no reference to the image (Table 1's § value for it counts image descriptions retrieved first). Its bottleneck is the write, as almost every poisoned record it writes is later used.

Each swap changes only retrieval, since the poison must be retrieved for the attack to succeed, these gains measure how often each encoder exposes the agent to it.

Paper Revision: The revised paper adds both encoder-swap studies to the appendix (F.12).

W3: Is the perceptual quality of the adversarial images evaluated quantitatively, e.g., with LPIPS or a human study?

Thank you for this suggestion. We adopt the ϵ=16/255\epsilon = 16/255 budget commonly used in existing transfer attacks such as M-Attack, FOA and M-Attack-V2 (Table 14). As requested, we now evaluate every attacked image on all four benchmarks with reference-based metrics, LPIPS (calibrated on human judgments; Zhang et al., 2018) and SSIM, and the no-reference quality model MUSIQ (Ke et al., 2021). We also ran VIVID at ϵ=8/255\epsilon = 8/255 and 4/2554/255.

Image (MemLens)LPIPS ↓SSIM ↑MUSIQ ↑
Source + JPEG-30 recompression0.0570.91843.1
M-Attack-V2, ϵ=16/255\epsilon = 16/2550.2110.77348.3
VIVID, ϵ=16/255\epsilon = 16/2550.2170.77846.6
VIVID, ϵ=8/255\epsilon = 8/2550.0990.90048.2
VIVID, ϵ=4/255\epsilon = 4/2550.0350.96650.8

Mean over the 47 goals, at the 224px resolution the attack uses. Source images score 52.2 on MUSIQ.

As shown, VIVID is no more perceptible than M-Attack-V2 at the same budget: its LPIPS is within 0.01 of M-Attack-V2's on all four benchmarks. These values match those published for ϵ=16/255\epsilon = 16/255 transfer attacks on VLMs: LPIPS of 0.209 for M-Attack and 0.217 for FOA, on different images. Against the source, the change is measurable (LPIPS 0.217, vs 0.057 for a JPEG-30 recompression), but without the source it does not read as degradation: MUSIQ places the attacked image between the source and its JPEG-30 copy on all four benchmarks.

Lowering the budget trades visibility for strength. At ϵ=4/255\epsilon = 4/255, LPIPS falls below the JPEG-30 copy and SSIM reaches 0.966, yet VIVID still transfers on 14.2% of draws, against 1.8 to 2.5% for noise of matched energy (Line 369). At ϵ=8/255\epsilon = 8/255 (LPIPS 0.099), it keeps about half of its ϵ=16/255\epsilon = 16/255 transfer (35.2% vs 63.6%).

Paper Revision: The revised version adds these metrics for all four benchmarks and the ϵ\epsilon sweep to the appendix (F.13).

References.

  • Xie et al. "Improving transferability of adversarial examples with input diversity." CVPR 2019.
  • Zhang et al. "The unreasonable effectiveness of deep features as a perceptual metric." CVPR 2018.
  • Ke et al. "Musiq: Multi-scale image quality transformer." ICCV 2021.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Paper Revisions

We would like to thank all reviewers again for their constructive feedback. We have uploaded a revised version of the paper and highlighted the major changes in blue for clarity. In summary:

  • We have added an end-to-end comparison with the strongest transfer baseline, M-Attack-V2, through the full memory pipeline (Sec. 7, Table 2), with per-backend and per-victim results in Appendix F.2.
  • We have clarified the contribution of delivery-aware optimization in Sec. 5.2, added it to M-Attack-V1, FOA and M-Attack-V2 in Appendix F.1 (Table 18), and named the corresponding ablation “M-Attack-V2 + delivery-aware optimization” (Fig. 9, Table 14).
  • We have stated our main finding, that the memory architecture sets how much of a visual poison survives and where it can be caught, in the abstract, the introduction and Sec. 7.
  • We have stated the population of each metric in Appendix E, added the clean-hit exclusion in Table 17, and added goal counts, retained counts and retention over ASR hits to Table 1 and Tables 22 to 25.
  • We have added a delayed-recall analysis in Appendix F.10 and a query-generalization study in Appendix F.11.
  • We have added image-retriever and text-embedder swap studies in Appendix F.12.
  • We have added perceptual-quality metrics and a perturbation-budget sweep in Appendix F.13, summarized in Sec. 6.

We appreciate the reviewers for the feedback that has contributed to significantly improving our paper, and we welcome further discussion.

Kind regards,

Authors

AuthorsOpenReview ↗

Follow-up and Clarifications Before Discussion Closure

We thank the reviewers for their insightful feedback. As the discussion period draws to a close, we would like to revisit our earlier responses to ensure we have thoroughly addressed the concerns and questions posed by the reviewers. If there are any additional questions or points requiring clarification, we are more than happy to provide further details.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.