acceptodds
Under review as a conference paper at ICLR 2027

MemoryMischief: Multimodal Memory Poisoning Attacks on Web Agents

Abstract

External memory has become a core component of modern web agents, enabling long-horizon reasoning through the retrieval of past experiences. However, this paradigm introduces a critical vulnerability: malicious content injected into memory can be persistently recalled and repeatedly influence agent behavior. In this work, we identify and systematically study multimodal memory poisoning, an overlooked yet practical attack surface in web-agent systems. We propose MemoryMischief, a unified black-box attack framework that poisons graph-structured external memory with coordinated text-image evidence. Our method consists of a two-stage design: (1) a trigger-conditioned retrieval attack that ensures high-probability recall of malicious memory, and (2) a post-retrieval attack induction that leverages adversarial perturbations and stealthy OCR injection to override the original user objective. Unlike prior attacks that operate on prompts or text-only memory, our approach enables persistent, reusable, and goal-agnostic attacks without modifying model parameters or re-optimizing malicious tasks. Experiments across multiple web-agent frameworks and vision-language models demonstrate that MemoryMischief achieves strong end-to-end attack success with minimal impact on benign performance, reaching up to 99.15% on GPT-5-family web agents, while transferring effectively across architectures and model scales.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, agent memory, multimodal agents, memory poisoning, adversarial machine learning, AI safety, web agents

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_4sUUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates multimodal memory poisoning against web agents. MemoryMischief combines a visual trigger optimized for malicious-memory retrieval with a reusable post-recall component using surrogate-based adversarial perturbations and OCR instructions. A separate memory component specifies the malicious goal. Experiments cover three agent frameworks, four VLM backbones, four sandbox attack categories, and two memory schemas. The paper reports high retrieval and action success, including up to 99.15% end-to-end success, together with component ablations and lightweight defense evaluations.

Critical strengths
  1. The study connects poisoned-memory retrieval to executable agent behavior. Separating retrieval success, conditional action success, and end-to-end success provides a useful evaluation framework. The deterministic sandbox checks also offer more concrete evidence than judging malicious textual responses alone.

  2. The component ablations show a meaningful multimodal contribution. In Table 4, adding the complete visual component increases end-to-end success from 73.59% for text-only poisoning to 93.97%. The evaluation across several agents and VLM backbones provides additional evidence that the observed vulnerability is not confined to one policy model.

Critical weaknesses
  1. The claimed black-box, memory-only threat model is not adequately reconciled with the method. Section 3.1 grants the attacker only memory-level access and excludes access to retriever internals. However, retrieval optimization explicitly uses the retrieval encoder’s embeddings and benign-memory embeddings, while attack activation requires a crafted trigger in the current webpage observation. The paper does not clearly explain how these capabilities are obtained under the stated restrictions. Figure 4 provides surrogate-to-target retrieval results, but does not establish the complete attack under an unknown retriever. This distinction matters because the headline success rates may depend on stronger access than the proposed practical threat model allows.

  2. The experimental protocol lacks details needed to assess the headline results and comparisons. The manuscript does not sufficiently specify task counts and splits, the default memory corpus size and retrieval depth, numerical trigger and perturbation settings, or the adaptations and tuning of the baselines. Attack success rates also lack trial counts and uncertainty estimates, although benign utility includes repeated-run statistics. Furthermore, all three agents use the authors’ shared memory interface, so the agent matrix provides limited evidence about independent memory implementations. These gaps make it difficult to verify whether the gains persist under matched poisoning budgets and realistic retrieval conditions.

Questions for authors
  1. What exact access is required during retrieval-stage optimization? Are the main results obtained using the victim retriever or a surrogate, and who controls insertion of the trigger into the current webpage?

  2. Can you provide the complete evaluation configuration, including task counts, train/test separation, memory size, retrieval depth, attack budgets, baseline adaptations, and confidence intervals for ASR-ra?

  3. Does a single injected memory remain effective across later episodes as benign memories accumulate, without reinsertion or re-optimization? This would materially strengthen the claimed persistence beyond reuse in fixed evaluation settings.

1 comment

Dear Reviewer,

Thank you again for your detailed and technically specific feedback.

We have made considerable efforts to address all three questions you raised. Beyond clarifying the threat model and surrogate-based optimization setting, we provided a more comprehensive experimental configuration, including task splits, memory size, retrieval settings, attack budgets, baseline adaptations, and repeated-run statistics. More importantly, we conducted an additional memory accumulation experiment, demonstrating that a single poisoned memory maintains substantial retrieval effectiveness as the memory corpus grows from 740 to 6418 vectors, without reinjection or trigger re-optimization.

We sincerely hope these additional clarifications and experiments address your concerns. As the rebuttal period is approaching its end, we would greatly appreciate any further feedback you may have.

Thank you very much for your time and consideration.

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

This paper studies multimodal memory poisoning against web agents. MemoryMischief first increases the chance of retrieving poisoned memory when a visual trigger appears. It then uses malicious text and visual signals to guide the agent toward an attacker’s goal. The method is evaluated across several agents, models, and controlled attack tasks.

Critical strengths
  1. Separating memory retrieval from action induction provides a clear way to understand the attack.
  2. The modular design allows different malicious goals to reuse the same attack components.
  3. Experiments cover several agents, model backbones, and attack goals. The ablations help explain the main components.
Critical weaknesses
  1. The attacker assumptions need stronger practical justification. Direct memory insertion and the appearance of a webpage trigger may require different forms of access.

  2. The black-box setting needs clearer qualification. Known-encoder optimization should be clearly distinguished from transfer to an unknown retrieval system.

  3. The agents share the same added memory interface. This supports controlled comparisons, but limits the evidence for generalization to independently designed memory systems.

  4. The downstream attacks use simplified sandbox tasks. These tests show that agents can follow malicious memory, but provide limited evidence about the difficulty of comparable attacks in real workflows.

Questions for authors
  1. What access does the attacker need to both poison memory and place the activation trigger?
  2. How much end-to-end attack success remains when the victim’s retrieval encoder and benign memory are unavailable?
1 comment

Dear Reviewer,

Thank you again for your careful review and constructive feedback.

We have made substantial efforts to address all four concerns you raised. In particular, we clarified the two independent attacker capabilities and the surrogate-based black-box setting, added end-to-end cross-encoder transfer experiments, provided evidence across graph-based and single-node/bundle memory schemas, and further explained the execution-grounded nature of our sandbox evaluation.

We sincerely hope that these clarifications and additional results have addressed your concerns. As the rebuttal discussion is nearing its end, we would greatly appreciate it if you could share any remaining questions or feedback.

Thank you very much for your time and consideration.

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

The paper investigates multimodal memory poisoning in web agents that use graph-structured external memory. It introduces a black-box attack framework (MemoryMischief) which has two stages: 1) trigger-conditioned retrieval optimization, which aims to make poisoned memory entries more likely to be retrieved when specific visual triggers (a patch) appear; and 2) post-retrieval composite visual attack, which combines adversarial perturbations with stealthy OCR injection to make the agent prioritize the recalled memory and steer its downstream actions. Experiments across multiple web-agent frameworks and VLMs show high attack success rates, minimal degradation in benign-task utility, and transferability across agent architectures and VLM backbones.

Critical strengths
  • Multimodal memory poisoning is an important and underexplored security problem for long-horizon web agents.

  • The method separates the attack into retrieval and post-retrieval induction stages. The evaluation covers multiple VLM backbones and web-agent frameworks, with solid analyses of attack success and benign-task utility.

  • The paper is well structured, well motivated, and clearly explains the approach.

Critical weaknesses
  • The role of graph structure in the attack could be clearer. The authors motivate the method around graph-structured memory, but it's unclear what graph links contribute to retrieving or propagating the poisoned evidence, and in which settings do they affect attack success? Is graph structure essential to the method, or primarily the memory implementation used in the main experiments?

  • The trigger optimization appears to use the retriever’s image encoder (\phi) and embeddings of existing memory items. Does optimization use the target retriever or a surrogate encoder? Does it require access to the actual memory embeddings? The setup needs further clarification.

Questions for authors

See Weaknesses section.

1 comment

Dear Reviewer,

Thank you again for your thoughtful review and positive assessment of our work.

We have provided detailed clarifications regarding the two concerns you raised, particularly the role of graph relations in component association and evidence propagation, as well as the use of a surrogate retriever rather than the target retriever during trigger optimization. We have also added cross-encoder end-to-end transfer experiments to further support our claims.

We hope these clarifications adequately address your concerns. As the discussion period is approaching its end, we would greatly appreciate any further feedback you may have, particularly if any issues remain unresolved.

Thank you again for your time and consideration.

Comments on the paper

1 comment

Dear Reviewers,

Authors’ kindly responded to address your concerns. Please check rebuttals and share your opinions with authors. Thanks for your contributions!

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.