acceptodds
Desk rejected at ICLR 2027

MASCSBench: A Benchmark for Evaluating Emergent Content Safety Risks in Multi-Agent Conversational Systems

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-agent LLM systems can distribute a restricted objective across agents and turns such that each local request appears mild while the assembled conversation recovers the harmful intent. Existing safety evaluation often scores isolated responses or a flat transcript, obscuring this composition failure. We introduce MASCSBench, a benchmark for distributed safety reconstruction with 3,195 seed topics, 10,311 validated decomposition chains, and 10,162 realized conversations across seven top-level risk categories and balanced 3-, 4-, and 5-turn settings. A key information boundary separates deployable evaluators, which receive only observable response or transcript content, from an annotation-conditioned Risk Propagation Graph (RPG) reference protocol, which additionally uses adjudicated reconstruction patterns for diagnostic analysis and is therefore not treated as a competing detector. On the 5-agent, 5-turn risk-bearing slice, response-level scoring recovers 12.63% of cases, while a long-context transcript classifier recovers 61.13%, showing that local scoring can substantially understate dialogue-level risk. The annotation-conditioned RPG reference recovers 68.52% and is used to localize when and across which roles evidence assembles, not to claim a deployable performance advantage. Validation includes benign controls, human studies, construction ablations, multiple backbones, and multiple orchestration frameworks. We release sanitized examples, metadata, scorer code, reproducibility scripts, and gated access for high-risk items at https://anonymous.4open.science/r/mascsbench.

Primary area:
datasets and benchmarks
Keywords:
Multi-Agent Systems, AI Safety, Safety Benchmarking, Multi-Turn Conversational Safety, Distributed Safety Reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aDwbOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces MASCSBench, a benchmark of 10,162 multi-agent conversations designed to study safety risks that emerge from combining individually benign responses. It shows that full-context evaluation detects substantially more risks than response-level evaluation (61.13% vs. 12.63%). The authors also introduce a Risk Propagation Graph (RPG) to analyze how risk-related information accumulates across agents and turns.

Critical strengths

The paper presents a relatively large benchmark with 10,162 multi-agent conversations, supported by human annotations, benign controls, and evaluations across multiple models and frameworks. The authors have also made an effort to document the dataset construction and validation process.

Critical weaknesses
  1. The paper is poorly written, with excessive defensive language (e.g., “not as a competing deployable detector or as evidence that graph structure alone improves detection” and “not a deployable baseline and not a claim that role-edge structure alone explains the context gap”). Similar statements are repeated throughout the paper. Citation formatting is also consistently problematic (e.g., “content production Hong et al. (2024); Liu et al. (2024)”), making the paper harder to follow.

  2. The paper does not establish why this is a distinct multi-agent safety problem. The benchmark is constructed to make individual responses appear safe, so the main performance gap is partly expected by design. Moreover, simple keyword filtering (60.50%) performs nearly as well as the full-context classifier (61.13%), raising doubts about whether the benchmark captures genuine semantic reconstruction.

Questions for authors
  1. What makes this a distinct multi-agent safety problem rather than a standard multi-turn safety problem?

  2. Given that keyword filtering (60.50%) performs nearly as well as the full-context classifier (61.13%), what evidence shows that the benchmark captures semantic reconstruction rather than keyword accumulation?

  3. How do the authors validate that recovering harmful intent from individually benign responses represents a genuine safety failure?

Reviewer_dMRWOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1(−1)
Read review
Paper summary

The paper studies "distributed safety reconstruction": a user spreads a restricted objective across multiple agents and turns so that each local request looks mild while the assembled role-labeled conversation recovers the harmful intent. It introduces MASCSBench, a benchmark of 3,195 seed topics, 10,311 validated decomposition chains, and 10,162 realized conversations over seven risk categories and 3/4/5-turn settings, built by a CoT-RL decomposition (reward 0.4R_cov+0.3R_coh−0.3R_risk) with a two-stage verification that combines rule/LLM local-mildness screening and human global-recoverability annotation (κ=0.84; retained only when ≥2 annotators agree). The central empirical claim rests on Table 5 (5-agent, 5-turn risk-bearing positive slice): response-level scoring (SA) recovers 12.63% of cases while a long-context transcript classifier (CC) recovers 61.13%, with keyword filtering (KF) at 60.50%. A separate, explicitly annotation-conditioned Risk Propagation Graph (RPG) reference produces DDRR/RADI/CRRA/GCRI diagnostics (68.52% pattern recovery), which the authors are careful to position as a benchmark-side audit tool rather than a deployable detector. Validation includes benign controls, two human studies, construction ablations, multiple backbones, and multiple orchestration frameworks. The paper is notable for its disciplined self-limitation and for a responsible-release framing (sanitized public examples, hashed items, gated/hosted access for raw high-risk prompts).

Critical strengths
  1. Disciplined separation of the deployable-evaluation claim from the annotation-conditioned RPG diagnostic, maintained consistently across every results table. The paper never lets benchmark-side supervision masquerade as a detector advantage (Sec. 1, Sec. 4.2, Table 4, Table 5 caption explicitly excludes RPG; Discussion "What it does not establish," p.8). This is the right methodology for a dual-use safety benchmark and makes the central claim interpretable and falsifiable rather than inflated.
  2. Responsible-release and anonymity hygiene are concrete, not boilerplate. The Ethics Statement, Appendix B and Appendix E specify sanitization, hashed item linkage, and gated/hosted access with logging and a use agreement; the held-out example (Fig. 5, Table 6) is a genuinely non-operational paraphrase; PDF metadata and the only external links are clean/anonymized. For a benchmark that decomposes harmful intents this materially lowers the misuse surface.
Critical weaknesses
  1. The headline SA-vs-CC gap is partly built in by construction and may be driven by lexical accumulation rather than semantic reconstruction, and no single-turn control establishes the "emergent" framing. By Eq. (1) risk-bearing items are selected so that max_i L_loc(u_i)=0 — no local piece is explicitly unsafe — which mechanically handicaps the response-local evaluator (SA=12.63%) relative to full-context CC (61.13%). More tellingly, keyword filtering reaches 60.50%, essentially tying CC, and the authors concede "this positive slice does not isolate semantic composition from accumulated lexical cues" (p.6) and "KF recall is consistent with lexical shortcuts" (p.8). There is also no direct single-turn request control showing that the composed dialogue elicits harm a direct ask would not, so the title's "Emergent … Risks" is broader than the evidence, which concerns the evaluation unit (detection), not harm production. Because this is the paper's primary empirical claim, the confound determines how much novelty the result carries: a pattern-blind role-aware judge, lexically matched hard negatives, and a single-turn baseline are needed to show the gap reflects distributed reconstruction rather than selection plus vocabulary.
  2. Human alignment of the reference diagnostic is weak on the harder study, and the local-mildness screen model is unspecified. On the 300-dialogue stratified study pooled GCRI–human correlation is r=0.422 (95% CI [0.32,0.51]), with several categories near zero (Disinformation & Manipulation 0.060, Cybersecurity 0.110, Self-harm & Suicide 0.151; Table 10), versus r=0.84 on the smaller 100-dialogue expert sample — suggesting the expert-sample agreement overstates validity. Separately, the "LLM safety screen" that decides local mildness (Sec. 4.1) is never named; if it shares a family with the SA evaluator or the generation backbones (which include GPT-4o and Claude-3.5, also used as decomposers), part of SA's low recall is guaranteed by construction. These bear directly on whether GCRI and the SA–CC contrast measure what the paper claims.
Questions for authors
  1. In Table 2 and Table 16, the seed-topic column sums to 3,189 and the chains column to 10,305, not the stated totals 3,195 and 10,311 (the realized-conversation column does sum to 10,162). What accounts for the difference of 6 in each — cross-category de-duplication, multi-label items, or a typo? Please correct or explain.
  2. Can you provide a single-turn direct-request control (refusal/detection rate for the same restricted goal posed in one turn), so the distributed setting is shown to elicit risk that direct asking does not?
  3. Can you add a pattern-blind role-aware judge and lexically matched hard negatives to separate semantic reconstruction from lexical accumulation (given KF≈CC) and from the Eq. (1) selection effect?
  4. Which model backs the LLM local-mildness screen, and is it the same family as the SA evaluator / CC / the generation backbones? How much of SA's 12.63% recall is attributable to the screening criterion itself?
  5. Given r=0.422 pooled and several per-category correlations below 0.16 (Table 10), how should GCRI's role/turn path diagnostics be trusted in exactly those context-dependent categories?
1 comment

Thank you again for your constructive feedback and for taking the time to consider our rebuttal.

As the discussion period is coming to an end, we would appreciate knowing whether our responses have addressed your main concerns. If so, we would be grateful if you could consider updating your score to reflect these clarifications.

Thank you for your time and consideration.

Reviewer_fY7qOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a benchmark that turns a given harm into distributed queries to test multi-agent LLM conversations. The taks is split across agents, but the idea is to detect harmful outputs given the concatenation of the outputs.

Critical strengths

The paper looks at a failure mode of existing LLM/agents and builds detector for the harmful outputs. They compare their method against a variety of baselines including rule based baselines and exiting approaches like Llamaguard.

Critical weaknesses

I find several weaknesses with the paper:

1/ I believe this attack is well known in literature and defenses already exist. The weak results (e.g. rule based approach performs at 60.5% already) makes the approach less appealing.

2/ The paper is a bit hard to follow sometimes -- particularly the text. E.g. content like "MASCSBench turns distributed safety reconstruction into a concrete benchmark task for multi-agent LLM conversations. Across 10,162 role-labeled dialogues, the principal result is deliberately narrow: isolated response scoring can miss risks visible in complete dialogue context. " can be better written.

Questions for authors

N/A

1 comment

Thank you for your time and constructive feedback. As the rebuttal period is approaching its deadline, we would greatly appreciate it if you could review our response and share any remaining comments or concerns. Your feedback would be very helpful in ensuring that we have adequately addressed your questions.

Thank you again for your time and consideration.

Comments on the paper

2 comments

I thank the authors for the response.

  • W2 & Q1: The authors acknowledge that the same problem can also occur in ordinary multi-turn conversations. I'm still not convinced that the paper shows what makes this problem different in the multi-agent setting.

  • W2 & Q2: The new experiments are helpful, but they don't really explain why keyword filtering (60.50%) performs almost the same as the full-context classifier (61.13%). I'm still not sure the benchmark is testing semantic reconstruction rather than just keyword accumulation.

  • Q3: The explanation is clearer now, but I'm still not convinced that the risks identified actually lead to harmful assistance.

Also questions for the response itself:

  • I'm curious why sentences like “This is a representation check, with no model-performance or learned-specialization claim” and “It does not quantify the semantic share of the original archive or establish a role-metadata gain” appear in the rebuttal. They sound very AI-generated and unnecessarily defensive to me.

  • The authors mentioned making presentation changes, such as fixing the citation formatting (all of the citations are wrong now), but I don't see any changes in the PDF.

  • The rebuttal itself is still hard to read, with too much complicated and defensive language. This was already one of my concerns in W1.

Therefore, I'm now more confident in my decision to give this paper a Clear Rejection.

Thank you for the direct feedback. Our previous reply was too indirect. The contribution is an evaluation task: determining which contributions from different agents jointly justify a safety judgment.

What the multi-agent setting adds. MASCSBench records the source, role, and order of contributions. Our new control makes their relevance concrete: an owner states permissions, an assembler identifies the source of records, and a coordinator states their destination. Evaluating the proposed transfer requires connecting these facts. The same inference can arise in ordinary multi-turn dialogue. We have not established a new failure mechanism caused by adding agents.

Semantic evidence. In the new authorization task, each paired message contains exactly the same words with the same frequencies. Roles, order, and lengths also match; which source the records come from changes. Given an explicit policy, GPT-OSS and Gemma each distinguish all 16 test pairs from eight contexts. In 13/16 risk cases, an executable check confirms that both models cite sufficient facts and that every cited fact is needed. The stricter test also requires correct judgments on every incomplete view and minimal evidence for the allowed counterpart; none passes (0/24 across development and test). All new model judgments are automatic.

Why the original KF and CC recalls are close. We checked 328 dialogues marked positive by the archived safety scores: both detect 128, KF alone 69, CC alone 77, and both miss 54. The similar scenario-average recalls conceal substantial disagreement. There is also a necessary correction: the recovered implementation reproduces 61.13% by aggregating local safety scores, not by independently classifying a full transcript. That number cannot establish semantic reconstruction. The word-matched experiment supplies separate evidence; lexical confounding in the original archive remains unresolved.

Harmful assistance. The 42-goal comparison, conducted under a shared policy requiring brief, non-operational answers, did not establish increased prohibited assistance in the multi-agent conditions. Our supported outcome is a safety judgment about the combined context. Increased harmful production remains unproven.

Presentation. The two sentences you quoted obscured these points. The citation and prose edits were only in the source; the PDF you reviewed had not been updated. We should have stated that status clearly.

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

Related papers

Discussion (0)

Sign in to comment.