acceptodds
Under review as a conference paper at ICLR 2027

MEMSHIELD: NEUTRALIZING COALITIONAL MEMORY POISONING IN LLM AGENTS VIA CAUSAL REPRESENTATION REPAIR

Abstract

Persistent memory empowers LLM agents but exposes them to joint memory-poisoning attacks that manipulate the retrieval representation space to induce cross-session harmful responses. While existing defenses effectively detect isolated poisoned records, they fall short against multi-memory attacks that distribute adversarial triggers across records. To bridge this research gap, we present , a causal memory-repair framework that retrieves and cleanses combined poisoned memories. We formulate memory repair as a causal representation intervention problem and introduce a multi-agent planner that determines optimal memory subsets to retain, mask, or quarantine. Specifically, the planner isolates the Minimal Triggering Set (MTS) driving the risk and the Minimal Blocking Set (MBS) required for risk suppression. To ensure repaired memory spaces cannot reactivate adversarial circuits, incorporates a closed-loop verification mechanism with pre-retrieval quarantine and post-retrieval re-gating. Across four attack families and six benchmarks, reduces the residual Attack Success Rate (ASR) to 0% with negligible impact on clean performance, while demonstrating strong cross-architecture transferability.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Joint Poisoning, Causal Graph, LLM Defense

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_UL1cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MEMSHIELD, a defense framework for persistent-memory poisoning in LLM agents, with a particular focus on attacks where multiple individually benign-looking memories jointly trigger malicious behavior. The method uses provenance/dependency graphs and counterfactual replay to identify Minimal Triggering Sets (MTS) and Minimal Blocking Sets (MBS), then quarantines selected memories and re-runs retrieval and planning to verify that the attack does not reappear through replacement memories.

Critical strengths
  1. The paper targets an important and underexplored threat model in which multiple memories jointly form an attack trigger that cannot be detected reliably through single-memory attribution.
  2. The combination of set-level counterfactual diagnosis with re-retrieval and re-planning is well motivated because quarantining one memory can change retrieval and expose alternative poisoned memories.
  3. The MTS/MBS formulation provides a useful conceptual distinction between memories sufficient to trigger harmful behavior and memories whose removal is sufficient to suppress it.
Critical weaknesses
  1. The empirical comparison is difficult to interpret because several baselines use different or adapted protocols, some method–dataset pairs are missing, and the paper explicitly avoids matched cross-protocol causal comparisons or significance testing.
  2. The main end-to-end evaluation uses a single protected model and a single reported seed, while the claimed cross-architecture transferability in the abstract is not supported by completed matched cross-model results in the current experimental export.
  3. Several advertised components, including the multi-agent debate planner, long-term experience reuse, and persistent cleanup, are not directly validated by matched end-to-end experiments, making the contribution of the full framework unclear.
  4. The framework appears computationally expensive, requiring potentially hundreds of replay calls and over one million tokens per diagnosis, yet the paper does not provide a matched latency or total-overhead comparison against baselines.
Questions for authors
  1. Can you provide matched experiments across multiple protected LLMs to support the claimed cross-architecture transferability?

  2. Why are the baseline comparisons not run under a common protocol and identical evaluation budget?

  3. How much do the multi-agent planner and long-term experience store actually improve safety or reduce diagnosis cost?

  4. What is the practical latency and token overhead of MEMSHIELD per triggered case compared with existing defenses?

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

The paper presents MemShield, a framework for defending large language model agents against coalitional memory poisoning attacks, where multiple individually innocuous memory records can jointly induce harmful behaviors across sessions. Unlike conventional defenses that primarily examine individual suspicious memories, MemShield formulates memory repair as a set-level causal intervention problem.

The framework constructs a provenance–dependency graph to identify potentially responsible memories and employs controlled counterfactual replay to characterize Minimal Triggering Sets (MTS) and Minimal Blocking Sets (MBS). It further introduces pre-retrieval quarantine and post-retrieval masking to account for retrieval-induced replacement effects, together with an optimization objective that balances security, legitimate task utility, and remediation costs. An optional multi-agent intervention planner and a long-term causal experience store are incorporated to facilitate diagnosis and subsequent repair.

Experiments cover several memory-poisoning attack settings and agent benchmarks, including StrategyQA, WebShop, ALFWorld, SafeOS, and AgentHarm. The reported results show zero observed attack successes across 1,347 scored attack instances, while clean-task performance varies across benchmarks. Ablation studies provide evidence for the usefulness of set-level interventions and counterfactual validation in selected settings. However, the evaluation does not yet establish the effectiveness of the complete framework, particularly its multi-agent planning, long-term experience reuse, and persistent memory cleanup components.

Critical strengths

S1: Addresses an important and underexplored security threat in persistent LLM agent memory. The paper focuses on coalitional memory poisoning, where multiple memory records jointly induce harmful agent behaviors despite appearing relatively benign when examined individually. This setting is practically relevant because persistent agent memories can accumulate information across sessions, creating attack opportunities that conventional record-level detection may overlook. Addressing this limitation represents a meaningful research direction for improving the security of memory-augmented LLM agents.

S2: Introduces a structured set-level causal diagnosis and intervention framework. The formulation of Minimal Triggering Sets (MTS) and Minimal Blocking Sets (MBS) provides a systematic way to reason about interactions among poisoned memories. By combining provenance–dependency graphs with controlled counterfactual replay, the proposed framework goes beyond individual-memory scoring and explicitly considers combinations of memories responsible for harmful behavior. This is a potentially useful methodological contribution, particularly for attacks involving redundant or interacting triggering paths.

Critical weaknesses

W1: The empirical evaluation does not sufficiently validate the complete proposed framework. Although the paper introduces several interconnected components, including multi-agent intervention planning, long-term causal experience reuse, and persistent provenance-based cleanup, the reported experiments primarily evaluate the core intervention pipeline. The optional multi-agent debate mechanism is not included in the complete ablation results, while the case studies do not demonstrate the full workflow. Consequently, the effectiveness of several claimed contributions remains largely unverified. This gap is particularly important because the paper presents these components as integral elements of its proposed framework rather than merely optional implementation details.

W2: The formal causal attribution and minimality claims are not adequately supported by empirical evidence. The identification of MTS and MBS relies on controlled subset interventions, candidate-domain closure, and sufficient search coverage. However, exhaustive evaluation becomes computationally expensive as the candidate set grows, and the proposed budget-constrained search may leave relevant combinations unexplored. Although the authors appropriately distinguish incomplete diagnosis from formally validated attribution, the experiments do not provide sufficient quantitative evidence regarding the frequency of successful MTS/MBS certification, candidate coverage, or minimality verification. Therefore, it remains unclear how reliably the central causal attribution mechanism operates in realistic large-scale memory environments.

Questions for authors

Q1. End-to-end validation of the complete framework. Could the authors clarify which components of MemShield were actually enabled in the main experiments? In particular, were the multi-agent intervention planner, long-term causal experience store, and persistent memory cleanup mechanism jointly activated? If not, can the authors provide an end-to-end evaluation demonstrating the effectiveness of the complete framework rather than only its core intervention components?

Q2. Reliability of MTS/MBS identification. How frequently does the proposed method successfully identify formally validated MTS and MBS under the configured replay budgets? Could the authors report the proportion of completed versus incomplete diagnoses, the average number of evaluated subsets, and the distribution of identified set sizes? These statistics are important for determining whether the proposed causal attribution formulation is practically achievable.

Q3. Fairness and statistical significance of baseline comparisons. Given the differences in benchmark adaptations, execution protocols, and scored sample sizes, can the authors provide additional comparisons under matched experimental settings? In particular, would the reported advantages remain significant when evaluated using identical attack instances, retrieval configurations, and statistical confidence intervals?

Q4. Computational efficiency and scalability. How does the computational cost of MemShield scale with the number of candidate memories and the complexity of their dependencies? Could the authors report average diagnosis latency, replay counts, token consumption, and end-to-end repair overhead per task, preferably compared with existing counterfactual attribution defenses? This would help assess the feasibility of the framework for real-world deployment.

Q5. Contribution of individual mechanisms. Can the authors provide additional controlled ablations for pre-retrieval quarantine, dynamic candidate expansion, multi-agent intervention planning, and historical experience reuse? In particular, how much improvement can be attributed specifically to dynamic retrieval-aware validation rather than simply applying more aggressive memory filtering?

Q6. Robustness under adaptive adversaries. How robust is MemShield against adaptive attackers who understand the defense mechanism and deliberately distribute poisoning triggers across a larger number of weakly related memories? Could such attackers exploit candidate selection thresholds, replay budgets, or incomplete provenance information to evade diagnosis? An adaptive evaluation would materially strengthen the paper's security claims.

Reviewer_Vyx1OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

MEMSHIELD addresses persistent memory poisoning in LLM agents through graph-guided candidate discovery, counterfactual subset replay, and authorized memory quarantine. It defines Minimal Triggering Sets (MTS) and Minimal Blocking Sets (MBS), accounts for retrieval replacements, and selects validated interventions using a cost-sensitive objective.

Critical strengths
  1. The paper addresses a relevant threat: interacting memories and alternative trigger paths can defeat individual-memory analysis, while retrieval replacements can undermine context-level mitigation.
  2. The framework carefully distinguishes behavioral responsibility, poisoning evidence, and remediation authority, as well as validated blocking effectiveness from certified minimality.
Critical weaknesses
  1. Unclear meaning of “causal representation repair.” The implemented interventions primarily mask context records or change retrieval eligibility rather than modify internal model representations.
  2. The experiments do not sufficiently demonstrate recovered MTS/MBS or verify their minimality (contribution 1). Ablation differences do not establish coalitional trigger structures.
  3. No matched ablation isolates pre-retrieval validation or adaptive candidate expansion. Aggregate ASR results therefore do not establish the contribution 2 of this central mechanism.
Questions for authors

How retention-loss, trust, and cleanup costs are computed, normalized, or calibrated?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.