acceptodds
Under review as a conference paper at ICLR 2027

BlameShift: Blame Integrity of Multi-Agent Failure Attribution under Ownership-Constrained Narrative Edits

Abstract

Multi-agent failure attribution relies on mixed logs combining trusted execution events with agent-authored messages. An agent responsible for a failure can exploit this interface by rewriting its own narrative to shift blame onto an innocent peer while execution records remain fixed. We define blame integrity as the invariance of an audit interface–judge mechanism to ownership-constrained narrative edits, precluding profitable deviations for blame-avoiding agents. On multi-agent benchmarks, ownership-legal accusations reliably steer flat LLM judges toward preselected victims over matched controls, an effect driven by peer accusations rather than text length, name priming, or factual fabrication. Vulnerability varies across judges, belonging to the audit mechanism rather than the text alone. Theoretically, evidence-independent style filters are either bypassable or own-blame inert, and empirically, filter-aware rewrites circumvent them. While narrative deletion trades off clean accuracy and citation verification limits steering, offline committees fail to Pareto-dominate their strongest constituent.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multi-agent systems, failure attribution, AI auditing, LLM-as-a-judge, adversarial robustness, accountability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_4mvsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines whether an agent responsible for a multi-agent failure can shift blame to a preselected innocent peer by editing only its own audit narrative, while the execution record and ground-truth responsibility remain fixed. To study this setting, the authors formalize blame integrity and evaluate ownership-constrained narrative edits on Who&When and TraceElephant traces. They also compare how different LLM judges respond to these edits and evaluate audit-interface defenses, including their effects on clean attribution.

Critical strengths

The paper presents a clear setting for targeted blame manipulation, where an agent can edit only its own narrative to shift blame onto a preselected innocent peer. Under Qwen-Max All-at-Once, named-F1 raises the targeted blame success rate over matched noise by +0.59 on HC, +0.38 on AG, and +0.41 on TE. Susceptibility varies considerably across judges, while the defense experiments report clean attribution accuracy and judge-query costs for HC.

Critical weaknesses

Gaming the Judge already demonstrated that rewriting agent reasoning can change judge verdicts while actions and observations remain fixed. Victim preselection and field-ownership constraints define a more specific attribution setting, but it remains unclear what new scientific insights arise specifically from these constraints. Table 5's “Fixed failure?” comparison is also ambiguous, as it does not clearly distinguish the execution outcome from the judge's verdict.

The claim that blame shifts arise from peer accusations rather than fabricated events remains insufficiently validated. In particular, Table 18 flags possible fabricated-event claims in some F1 rationales, and no human validation is reported for the subset unflagged by both automatic auditors.

The supplementary evaluations raise concerns about sample handling and comparability with the main evaluation. Table 10's caption refers to all 50 AG traces, although only 47 have admissible victims in the main targeted evaluation. The treatment of the remaining three cases in the targeted-lift calculation is not specified. Separately, Table 16 scores the blind attack against a writer-selected victim, limiting direct comparison with the preassigned-target setting.

The theoretical treatment does not fully account for the implemented lexical filter. Definition 3(ii) retains or drops entire narratives, whereas Appendix D removes selected lines. This difference raises questions about the applicability of Proposition 1 and Corollary 1 to the implemented filter.

The table captions are consistently placed below the tables, contrary to the ICLR 2027 formatting guidelines.

Questions for authors
  1. What new finding depends specifically on victim preselection and field ownership beyond Gaming the Judge? In Table 5, does “Fixed failure?” refer to the execution outcome or the judge's verdict?

  2. Please provide human validation for a sample of the Table 18 rationales that neither automatic auditor flagged, or revise the abstract's fabrication claim to reflect what the automated screening actually supports.

  3. Table 10 states that both axes use all 50 AG traces, whereas Table 1 excludes three traces without admissible victims from the main evaluation. How is targeted lift defined for these three traces?

  4. Appendix N dismisses peer-name filtering, while Table 15 reports detection results without evaluating this specific filtering defense. Why would removing peer names from agent-written narratives prevent attribution using trusted records? Has the effect on clean attribution accuracy or susceptibility to adaptive attacks been evaluated?

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

This paper focus on failure attribution in multi-agent settings. Specifically, the authors suppose one agent who did harmful operation might edit its own rationale to shift blame to other agents even though the execution logs were immutable. Then, the author formalize blame integrity as an property of audit interface and judge pair (ϕ,J)(\phi, J), showcase the impact for the final blame judge with misleading edited rationale. However, the results demonstrate only Qwen-Max was affected substantially.

Critical strengths

S1. The author give a clear threat model and it is possible that the dishonest agent can change the narratives given immutable execution logs.

S2. The ablation controls are reasonable with crafted baselines.

Critical weaknesses

W1. Lack of generalizability. The empirical vulnerability is located in a single model family plus a specific prompting style, flat All-at-Once Qwen-Max. For other frontier models including GPT-5.6, Claude-Opus, Gemini-2.5-Pro, the impact is marginal. Therefore, it is hard to derive a real failure of multi-agent attribution. Besides, there is no analysis on why Qwen-Max fails and others do not.

W2. The evidence is not solid and not enough to support author's conclusion given a small sample size (n≈50n \approx 50 per split) and the confidence intervals cover zero. Besides, the theoretical part does not offer useful information, e.g., Proposition 1 and Corollary 2 follow by construction.

Questions for authors

Q1. Given that most tested models including GPT-5.6 and Claude-Opus are naturally robust to the attack (lifts near zero), is this vulnerability truly a phenomenon worth investigating? Or is it merely an "artifact" resulting from the specific way Qwen-Max follows instructions?

Q2. In evaluation, citation verification can exactly filter fake narratives out but cannot distinguish the case with real but misleading statement. Beyond existing test, can the authors consider an adaptive attack that introduce selective disclosure of true receipts against the citation-verified judge?

Reviewer_9eX2OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces blame integrity, where an agent responsible for a failure attempts to shift blame onto an innocent peer by modifying its own narrative while keeping execution records unchanged. The authors use LLM-generated, self-exonerating rationales appended to attacker-owned messages and evaluate their influence on LLM-based failure attribution using Who&When and TraceElephant. They also formalize blame integrity and investigate several attribution interfaces and defenses.

Critical strengths
  • Interesting problem formulation: The paper raises a relevant integrity concern in multi-agent accountability, where an agent can manipulate its own narrative to influence downstream failure attribution without modifying execution evidence.
  • Controlled evaluation: The authors evaluate the attack across multiple datasets and attribution mechanisms, including matched controls and defensive interfaces, showing that certain LLM judges are susceptible to blame-shifting narratives.
Critical weaknesses
  • Unclear practical threat model and attack feasibility: The main attack assumes that the responsible agent can modify its final writable message after execution, with knowledge of the true responsible agent and a preselected innocent victim. While the authors explore removing this knowledge and appending a new message, the paper does not sufficiently establish how such post-execution narratives would be generated or admitted in realistic multi-agent auditing systems. It is also unclear why conflicting agent narratives would be treated as authoritative rather than evaluated against immutable execution records.
  • Limited generalizability and questionable practical significance: The attack largely consists of appending an LLM-generated accusation and self-exonerating rationale to an agent's message. Its effectiveness is highly dependent on the attribution interface and judge, with several evaluated models showing little or no significant susceptibility. This raises questions about whether the paper demonstrates a general integrity vulnerability or primarily the susceptibility of a particular LLM-based attribution setup. The writing and presentation further make it difficult to distinguish the fundamental security issue from the specifics of the experimental construction.
Questions for authors
  1. Threat model: In a realistic deployed multi-agent system, how does the responsible agent obtain the opportunity to modify or append a post-execution rationale, and what guarantees that this content becomes part of the audit evidence considered by the judge?
  2. Conflicting evidence: Since the attacker cannot modify another agent's messages or the trusted execution records, how does the proposed attack differ from an agent simply making an unsubstantiated accusation? Why should such a statement meaningfully alter attribution in an appropriately designed auditing system?
  3. Generalizability: Given that several evaluated judges are resistant to the attack, what evidence supports the claim that blame shifting is a broader multi-agent integrity problem rather than a weakness of specific LLM judges and flat attribution interfaces?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.