acceptodds
Under review as a conference paper at ICLR 2027

Judging Before Propagation: Task-Graph Contextual Integrity for Multi-Agent Privacy

10 views

Abstract

Multi-agent LLM systems increasingly rely on persistent memory and inter-agent communication. These capabilities create privacy risks that extend beyond individual interactions: a release appropriate in isolation may become inappropriate when combined with information previously disclosed to a recipient or propagated to downstream agents along the planned workflow. Existing Contextual Integrity (CI) mechanisms for agentic systems typically rely on local judgments that assess each disclosure in isolation, with or without regard to the recipient's history, but do not account for how information propagates downstream. However, compliance at the current boundary does not guarantee appropriateness across the workflow. We introduce , to our knowledge the first task-graph-native formalization of Contextual Integrity for multi-agent systems. TG-CI represents a workflow as a and extends CI from atomic transfers to stateful, graph-level information flows by jointly considering an item's intended propagation scope and recipients' evolving disclosure states. We operationalize TG-CI through *propagation-scoped authorization* under two complementary policies: , which requires admissibility across the complete downstream scope, and , which confines each item to its maximal admissible region using lightweight frontier gates. We further introduce , a benchmark of 144 cases across four scenarios that jointly vary MATG topology, disclosure history, compromised-agent placement, and item-level privacy policies. In our primary experimental configuration, TG-CI reduces the mean unauthorized disclosure count by 85.8-94.1% relative to the Local-History CI baseline, while retaining 93.6-96.1% of the baseline's task accuracy, with the two enforcement policies providing complementary privacy-cost-utility trade-offs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multi-Agent Systems, Contextual Integrity, Privacy Leakage, LLM Agents, Privacy Benchmarks

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_Dh2QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that existing Contextual Integrity (CI) defenses for LLM agents judge each disclosure locally, possibly with the recipient's history, and so miss "transitive" leaks, where a disclosure that is fine at the first hop later travels to an inappropriate downstream agent. It proposes Task-Graph Contextual Integrity (TG-CI). The workflow is modeled as a fixed Multi-Agent Task Graph (MATG). A memory item is judged together with its intended propagation scope P (a rooted subgraph), and it is admissible only if every atomic CI flow on every edge of P is admissible given the recipient's disclosure state D_t (Eq. 11). Two enforcement policies are given. Full-Scope releases an item only if the entire downstream closure of the requester is admissible. Auto-Region releases the item over the maximal rooted admissible region and installs "frontier" filters on the edges where the region ends. These filters use MiniLM similarity plus DeBERTa NLI to strip matching claims from inter-agent messages. The authors also release MAPLE-Bench: 144 synthetic cases (48 workflows x 3 topologies) in four domains, each with 4 agents, 30 memory items, an item-recipient policy matrix, 2 cumulative fact-pair rules, and 5 multiple-choice questions. In the main HAIKU-LUNA setting, averaged over scenarios, the mean unauthorized disclosure count is 15.84 (No Protection), 10.47 (Local-History CI), 6.11 (Online Atomic CI), 1.49 (Auto-Region) and 0.62 (Full-Scope) (Section 5.3). Utility stays at 93.6-96.1% of Local-History CI. The cost is a higher judge-token count.

Critical strengths
  1. Relevant, well-motivated problem: Privacy leaks across multi-hop agent workflows (direct, cumulative and transitive, Figure 1) are a real and growing concern. Separating transitive disclosure as its own failure mode is a useful framing.
  2. Conceptual separation: The paper keeps three things apart: the MATG, the per-item propagation scope, and the recipient disclosure state. Atomic five-parameter CI is kept at each edge (Eqs. 8-11). The conservative-state argument (Appendix A) and the transaction-local tentative state for items within one response (Section 4).
  3. Practical design: Auto-Region with item-specific frontier gates is a sensible engineering idea. It does costly semantic inspection only on edges where a protected item must be removed. Messages to non-MATG neighbors are blocked outright.
  4. Reasonable reporting breadth: The paper decomposes leaks into direct, cumulative and transitive parts (Figure 3). It reports compromised-agent exposure as distinct items rather than events (Figure 4), breaks results down by topology (Figure 5), and covers three agent-judge model pairings (Table 2, Appendix C). It provides the judge rubric and schema (Appendix F) and promises code and data.
Critical weaknesses

Even with the strengths the work has the following limitations:

  1. Limited technical novelty: TG-CI is the conjunction of per-edge CI checks over a reachable subgraph (Eq. 11). Auto-Region is a reachability expansion that stops at the first inadmissible edge (Eq. 13). Binding a label or authorized scope to data and enforcing it at downstream boundaries is the central idea of classic information-flow control and taint tracking. The paper itself cites agent IFC systems that do this (Costa et al., 2025; Debenedetti et al., 2025). The paper does not explain what TG-CI adds over "IFC labels whose allowed sinks are decided by an LLM CI judge", and it does not compare against such systems. There are no theorems or guarantees beyond the short monotonicity argument in Appendix A. The "first task-graph-native formalization" claim therefore rests mostly on notation. Table 1, which marks only TG-CI as covering transitive disclosure, leaves out IFC-based agent defenses.

  2. Possible circularity between the judge's inputs and the ground truth: The ground truth is the policy matrix Y (Y_{i,v} = 1 iff v may receive m_i) written by the benchmark authors (Section 5.1). The judge receives "applicable norms" and a recipient task for every edge (Appendix F, Listing 2). The paper does not say how these norms are produced per case, or whether they are derived from, or written alongside, Y. If the norms encode Y nearly directly, the judge becomes close to an oracle lookup, and the large gains mostly reflect the fact that TG-CI queries that lookup on every downstream edge. The paper also never measures the judge's own accuracy (precision/recall of Adm_t against Y), so we cannot tell how much of the result comes from the graph mechanism and how much from the judge.

  3. The comparison with Online Atomic CI is unclear and may be confounded: Online Atomic CI "applies the same edge-level atomic CI used by TG-CI independently at every realized memory and inter-agent transfer" (Section 5.2). If every realized transfer is checked with the same judge, it is not clear why transitive leaks remain its "dominant residual source" (Section 5.3).

  4. The cost argument does not fit the results: Contribution (2) says the design "avoids expensive and repeated contextual re-evaluation along downstream routes". Yet both TG-CI modes use more judge tokens than Online Atomic CI in every scenario. In Healthcare, for example, the counts are 117.3K (Online Atomic CI), 126.8K (Auto-Region) and 130.6K (Full-Scope) (Figure 3a). The cost metric also leaves out MiniLM/DeBERTa inference (Section 5.2). No latency or wall-clock numbers are given. The efficiency claim is not supported as stated.

  5. The utility cost is understated: The abstract and Section 5.5 stress 93.6-96.1% retention, but this holds only for HAIKU-LUNA. Under HAIKU-HAIKU, Full-Scope falls to 62.22%, compared with 78.06% for Local-History CI and 81.04% for No Protection, which is about 77% of No Protection utility (Table 2). The paper should report, and analyze, false blocking (task-necessary items wrongly withheld), using the "7 necessary" items annotated in Table 3.

  6. The threat model is not instantiated in the experiments: Section 2 describes compromised agents that "deviate arbitrarily", issue overbroad memory requests, misrepresent task necessity, and transform content. Section 5 never says how compromised agents are implemented: what prompts, what attack strategies, and whether they adapt to the defense. Two points matter in particular. (a) The frontier gate relies on fixed thresholds (cosine 0.35, NLI 0.70; Appendix D), and an adaptive adversary could paraphrase or combine items to evade it. (b) Appendix E admits that semantic transformations by intermediate agents are not modeled. Without adaptive attacks, the "compromised-agent exposure" results (Figure 4) say little about robustness.

  7. Strong assumptions limit applicability: The MATG must be fixed and known before execution, with no dynamic replanning (Section 4). Each node's subtask and each edge's transmission rule are trusted. Memory must be split into clean <subject, attribute, value> items. In practice, task graphs in agent systems are often emergent, and an attacker who can influence planning or subtask descriptions can widen the authorized scope. None of these assumptions are stress-tested.

  8. The benchmark is narrow and designed around the method: The benchmark has 4 agents per case, shallow graphs, 2 cumulative rules per case, and items, policies and questions synthesized with generative AI (AI Use Statement). It was built to exhibit exactly the three failure modes TG-CI targets, which favors the method. No existing benchmark is used for evaluation, although several are discussed (AgentLeak, PiSAs, AgentSCOPE, MAMA; Section 5.1). Results on at least one external benchmark would greatly increase confidence.

  9. Missing baselines: OCELOT and CAMP are skipped because no public code was available (Section 5.2). Both are described in papers and could be reimplemented in simplified form. Other natural baselines are also missing: a static IFC/label baseline that enforces Y directly, AirGapAgent-style minimization, and a "downstream-aware prompt" baseline where the local judge sees the downstream task graph in one call.

Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Author Response (1/2): Contribution, Norm Provenance, Leakage Analysis, and Cost Clarification

W1. Contribution and relationship to IFC.

We thank the reviewer for the opportunity to clarify the contribution. TG-CI contributes a contextual authorization formulation that makes downstream propagation and recipients’ evolving disclosure states part of the current privacy decision. For example, Full-Scope can withhold an otherwise locally admissible release because its downstream scope is inadmissible. TG-CI integrates this reasoning with flow enforcement, addressing both the determination and enforcement of disclosure permissions.

Our additional FIDES-based static reader-set baseline, supplied with oracle per-item permissions from Y, still incurs 1.688 unauthorized disclosures per run from combinations of individually permitted fragments. This exposes a limitation of static per-item permissions alone and motivates TG-CI’s history-conditioned judgments. We will revise the introduction to acknowledge IFC’s established propagation controls and distinguish them from TG-CI’s contextual authorization contribution.

W2. Norm provenance and atomic-judge evaluation.

Our norms were authored from common-sense privacy expectations and relevant legal principles to model organizational privacy requirements. They were neither copied nor derived from Y (specific item–recipient permission labels), which is excluded from runtime judgment. For example, Appendix F supplies: “Disclose only information appropriate for arranging the requested service, and no more revealing than needed.” Norms and annotations necessarily express the intended privacy objective, but this policy consistency is distinct from supplying the answers. Moreover, Online Atomic CI receives the same norm inputs, as discussed in response to reviewer KAfd - W4&Q3, so TG-CI does not obtain privileged access to evaluation labels.

We additionally evaluated Full-Scope's atomic judgments against YY, treating permit as the positive class:

Agent / JudgePrecisionRecall
Claude / GPT97.58%93.30%
Claude / Claude90.47%78.87%
GPT / GPT97.76%91.03%

These results demonstrate substantial but imperfect agreement with benchmark permissions. Under Claude/GPT, among 79 candidate releases incorrectly permitted by the immediate atomic judgment, Full-Scope prevented 24 (30.4%) through downstream CI vetoes. These observations provide direct evidence that prospective downstream adjudication can prevent some unauthorized releases that would otherwise pass immediate atomic CI checks, demonstrating additional protection beyond immediate judge accuracy.

W3. Leakage breakdown & analysis.

We thank the reviewer for identifying this potential confound. We conducted an additional channel-level audit and evaluated Online Atomic CI with the same MATG route guard used by TG-CI (see detailed results in our response to KAfd - W3&Q1). Under Online Atomic CI, 78.9% of unauthorized disclosures occur through off-graph messages (disclosures along sender–recipient routes absent from the approved MATG), despite per-transfer CI checking. Adding the TG-CI route guard eliminates off-graph leakage, demonstrating the value of explicit route restriction as an integral component of TG-CI’s enforcement.

Importantly, Full-Scope provides substantial additional protection even after controlling for route enforcement. Compared with route-guarded Online Atomic CI, Full-Scope reduces 51.7% unauthorized disclosures, including a 56.0% reduction in memory-release leakage and a 42.7% reduction in on-graph/final-output leakage, with only a 0.69-percentage-point utility decrease. Auto-Region instead offers a different privacy–utility trade-off: it preserving more utility than the guarded atomic baseline (78.61%) at the cost of slightly higher residual leakage. We will incorporate the analysis into the revised manuscript.

W4. Computational efficiency and token cost.

We thank the reviewer for highlighting this ambiguity. Avoiding repeated per-edge CI judgments does not necessarily reduce overall execution cost. TG-CI reuses prospective authorizations to avoid redundant downstream adjudication, but total cost also depends on the number of prospective edges evaluated, the natural information-exchange rate of the workflow, and changes in agents’ retrieval and communication behavior. For example, stricter source-level blocking may trigger additional memory requests. Accordingly, our results in Sec. 5 explicitly demonstrate a privacy–token-cost trade-off, with TG-CI achieving stronger privacy protection at moderately higher CI-judge token cost. We will refine our efficiency claim to emphasize the reduction of repeated downstream judgments without implying lower overall computational cost.

AuthorsOpenReview ↗

Author Response (2/2): Additional Experimental Analysis and Clarification

W5. Utility retention and false-blocking analysis.

We thank the reviewer for highlighting. The retention reported in the abstract corresponds to our designated primary Haiku–Luna setting. We additionally measured Full-Scope task-necessary item withholding rates: 18.45% (Claude/GPT), 34.16% (Claude/Claude), and 17.07% (GPT/GPT). With Claude agents fixed, the Claude judge exhibits substantially higher false blocking and lower Full-Scope utility, despite nearly identical unprotected task accuracy. These findings suggest that judge reliability is an important factor in the privacy–utility trade-off, motivating improvements in judge capability and calibration. We will incorporate this analysis alongside the W2 judge evaluation in the revision.

W6. Threat-model instantiation and adversarial robustness.

We conducted a prompt-induced paraphrasing stress test, where mean unauthorized disclosures increased from 1.488 to 1.727 for Auto-Region and 0.624 to 0.637 for Full-Scope. Both variants retain substantial containment, with Full-Scope showing greater resilience. TG-CI's conservative disclosure-state accounting tracks authorized items independently of wording, while Full-Scope additionally prevents inappropriate information from entering the workflow. However, we acknowledge fixed MiniLM/NLI thresholds remain vulnerable to semantic evasion, and this preliminary test does not establish robustness against fully adaptive adversaries.

Our focus is TG-CI's formalization, framework, and proof-of-concept validation, while systematic robustness evaluation and enforcement are for future work. We will include these results and clarify the attack scope and limitations in the revised manuscript.

W7. Fixed workflow and structured memory assumptions.

We acknowledge these assumptions and clarify their scope. As stated in the paper, supporting dynamic replanning is left to future work, as it would require an additional trusted graph-monitoring mechanism to detect changes and update authorizations. Our present work instead focuses on the formulation, enforcement framework, and empirical validation of TG-CI under a predefined, trusted MATG. Regarding structured memory, the <subject, attribute, value> representation does not necessarily require manually curated records: such information can often be extracted from contextual descriptions or existing memory using information-extraction techniques. We will clarify these assumptions and limitations in the revision.

W8. Benchmark representativeness and external validation.

Our benchmark is designed to jointly evaluate direct, cumulative, and transitive privacy leakage in task-driven multi-agent workflows. MAMA primarily studies topology-dependent private-memory extraction, making its evaluation focus different from ours. Nevertheless, we adapted the MAMA dataset to our framework to examine whether our findings generalize beyond our constructed scenarios. Mean unauthorized disclosures per run on 150 runs:

MethodMemoryOn-graphFinalOff-graphTotal ↓Utility ↑
No Protection3.2404.5730.9932.10010.90771.85%
Local-History CI0.0000.0930.0530.1800.32767.75%
Online Atomic CI0.0000.0200.0000.0730.09369.95%
TG-CI Auto-Region0.0000.0400.0200.0000.06070.08%
TG-CI Full-Scope0.0000.0000.0000.0000.00069.00%

The results follow the similar trend as our original benchmark. Auto-Region reduces leakage by 35.7% relative to Online Atomic CI with comparable utility, while Full-Scope achieves zero observed violations. Utility reductions relative to No Protection remain modest; as explained in our response to KAfd–Q5, less privacy filtering does not guarantee that agents retrieve or correctly use the required information. These findings provide complementary evidence of TG-CI's privacy–utility benefits, although the two benchmarks are not directly comparable due to their different evaluation objectives. We will include the results and adaptation details in the revision.

W9. Comparison with additional privacy defenses.

We thank the reviewer for suggesting additional baselines. Our IFC comparison is addressed in W1. We additionally implemented an AirGapAgent-style data minimization baseline within our multi-agent framework. Under Claude/GPT, it incurred 15.652 unauthorized disclosures per run (5.937 memory, 8.413 off-graph, and 1.302 on-graph/final), consuming 34.1K defense LLM tokens.

We skip OCELOT and CAMP as OCELOT relies on a specialized fine-tuned defender model whose checkpoint was unavailable for reproduction, while CAMP targets cumulative PII exposure in multi-turn conversations through retroactive masking, rather than recipient-specific information flows across agents. We will incorporate the AirGapAgent-style comparison in the revised manuscript.

Thank you again for your valuable comments and constructive suggestions. We have carefully addressed your concerns in our rebuttal and provided additional clarifications where necessary. As the discussion period is limited, we would be most grateful if you could find a moment to look over our response and let us know whether our explanations have adequately addressed your questions. Should any concerns remain, we would be very happy to provide further clarification. Since our experiments take considerable time to complete, an early reply would be of great help in allowing us to prepare any additional results you may wish to see within the remaining discussion period. We sincerely appreciate your time and effort in reviewing our paper.

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

The paper proposes Task-Graph Contextual Integrity (TG-CI). The setting is a multi-agent workflow that is planned ahead of time as a fixed Multi-Agent Task Graph (MATG). The idea is that a memory item should be released to an agent only if an LLM-judged CI predicate holds on every edge of the item's downstream "propagation scope", and not just on the first hop. The authors offer two ways of enforcing this. In Full-Scope mode, the item is blocked if any downstream edge fails the check. In Auto-Region mode, the system computes the largest admissible subgraph rooted at the recipient and places "frontier" filters at its boundary. These filters use MiniLM similarity and DeBERTa NLI to strip the item from any message that tries to cross the frontier.

The authors also build MAPLE-Bench, which has 144 synthetic cases (48 workflows, each instantiated in 3 topologies with 4 agents).

Critical strengths
  • The problem is real, and the paper motivates it well. A release can look perfectly fine at the first hop and still leak transitively further down the planned workflow, and the three violation modes in Figure 1 are a useful way to think about this.
  • I like that the binary judge is kept separate from the enforcement mechanism (choosing the scope, placing the frontier gate). It makes the design easy to follow.
  • Doing the semantic filtering only at the frontier, instead of on every edge, is a sensible engineering choice.
  • The benchmark labels direct, cumulative, and transitive violations separately, and since the data is fully synthetic it avoids any privacy concerns of its own.
  • The prompts, the judge rubric, and the schema are all in Appendix F, which helps a lot with reproducibility.
Critical weaknesses
  1. There is not much to the formalism. Eq. (11) is just the AND of per-edge CI judgments over a subgraph. Full-Scope amounts to "check every reachable edge", and Auto-Region is a BFS that stops at the first failing edge. Nothing is proved beyond the fact that P∗P^* satisfies TG-CI by construction. In other words, the claim of the "first task-graph-native CI formalism" is a claim about notation rather than a theoretical result. Note also that the cumulative component (the ⊕\oplus operation on the space HH) is never actually defined. In practice, it is a list of item names handed to the LLM.

  2. The most natural baseline is missing, namely information-flow control. Keeping a secret from spreading transitively is exactly what IFC and taint/label propagation were designed for. The paper even cites CaMeL (Debenedetti et al.) and FIDES (Costa et al.) in the threat model, but never compares against them. I suspect that a simple label-propagation baseline using the benchmark's own policy matrix YY would remove most of the transitive violations at close to zero LLM cost.

  3. The baselines are weak, and the authors built all of them. Local-History CI only checks the memory-to-agent edge, so by design it cannot catch transitive leaks, and beating it on transitive leakage is guaranteed. The comparison that actually tells me something is against Online Atomic CI, which runs the same atomic judge at every transfer that really happens. Unfortunately, the paper never explains why Online Atomic CI still leaks 6.1 items on average. If the reason is that compromised agents paraphrase content to get past the per-hop checks, then the same paraphrasing should also get past TG-CI's frontier gate. I need to see where Online Atomic CI fails before I can accept the central claim.

  4. The method and the benchmark look circular to me. The ground-truth matrix YY is described as a "task-conditioned item-recipient policy", and the TG-CI judge is given exactly the recipient's task and the edge's transmission rule. So the benchmark is built such that knowing each downstream agent's task is enough to decide whether a release is admissible. That is precisely the information TG-CI looks ahead to collect, and precisely what the baselines do not get. The paper does not discuss whether YY was annotated independently of how the method was designed.

Questions for authors
  1. Can you break down the transitive violations that Online Atomic CI still lets through? Are they due to judge mistakes, to compromised agents transforming the content, or to something else? And why would TG-CI's gate not fail in the same way?
  2. How does TG-CI compare to a deterministic label-propagation (IFC) baseline that uses the same YY?
  3. Was YY annotated independently of the TG-CI judge prompt? Who did the annotation, and what was the inter-annotator agreement?
  4. Can you report results against an adaptive compromised agent that paraphrases, splits, or encodes the protected values in order to evade the MiniLM/NLI gate?
  5. Why does Local-History CI have better utility than No Protection under Haiku-Luna?
3 comments
AuthorsOpenReview ↗

Author Response (1/2): Formalism, Relationship to IFC, and Leakage breakdown

W1. Formal contribution and novelty.

We thank the reviewer for the opportunity to clarify the contribution. TG-CI extends CI’s information-flow object and decision scope from an atomic transfer to a stateful, graph-scoped disclosure for multi-agent systems. It deliberately preserves CI’s five factors at each edge, while making the intended propagation scope and recipients’ evolving disclosure states explicit components of the graph-level decision (Sec. 3, Eqs. (2)–(11)). This changes whether and where information is authorized to flow. Equation (11) composes atomic CI judgments to evaluate this broader information-flow object. TG-CI’s contribution is the contextual policy formulation that makes intended downstream propagation and recipient disclosure state part of the current authorization decision, together with its operationalization and evaluation.

We agree that H\mathcal{H} and ⊕\oplus require explicit definitions. Let A\mathcal{A} be the universe of privacy-relevant subject–attribute–content records ⟨d,a,x⟩\langle d,a,x\rangle, including generalized disclosures, and H=2A\mathcal{H}=2^{\mathcal{A}}. The state Dt(v)∈HD_t(v)\in\mathcal{H} records information actually delivered to recipient vv through monitored channels before step tt. For H,Δ∈HH,\Delta\in\mathcal{H}, define H⊕Δ:=H∪ΔH\oplus\Delta:=H\cup\Delta, yielding Dt+1(v)=Dt(v)⊕Δt(v)D_{t+1}(v)=D_t(v)\oplus\Delta_t(v), where Δt(v)\Delta_t(v) denotes information actually delivered after generalization or filtering. Thus, ⊕\oplus accumulates disclosed records, while the history-conditioned CI predicate assesses their cumulative appropriateness.

We will revise the presentation to make the contribution clearer and add the definitions to the revised manuscript.

W2&Q2. Contextual authorization and the FIDES comparison.

FIDES (Costa et al., 2025) enforces specified information-flow policies by tracking and propagating confidentiality and integrity labels. TG-CI additionally determines disclosure admissibility from the recipient’s task context, accumulated disclosures, and prospective downstream MATG, and uses the resulting propagation-scope authorization to govern source admission and subsequent communication. Our LLM-token overhead arises primarily from these contextual judgments. Comparing this cost with enforcing already-supplied labels omits the policy-determination problem that TG-CI addresses.

To address the requested comparison, we evaluated FIDES based label-based IFC baseline with oracle per-item reader permissions from the benchmark matrix YY, which TG-CI does not receive during execution. This configuration still produced 1.688 unauthorized disclosures per run on average. The residual cumulative violations involve individually permitted fragments that jointly reveal sensitive information, a restriction that static per-item reader sets alone do not capture.

We will include these results and clarify the relationship among CI, TG-CI, and IFC in the revised manuscript.

W3&Q1. Leakage breakdown and TG-CI’s contribution.

We analyze leakage by delivery channel in the Claude–GPT agent–judge setting. We also include Online Atomic CI with the MATG route guard used by TG-CI.

MethodMemory ReleaseOff-graphOn-graph + final answerTotalUtility
Online Atomic CI0.8894.8390.4026.13080.69%
Online Atomic CI + TG-CI route guard0.87700.4151.29278.61%
TG-CI Auto-Region0.86400.6241.48880.00%
TG-CI Full-Scope0.38600.2380.62477.92%

Our audit locates 78.9% of Online Atomic CI’s leakage in off-graph messages (disclosures along sender–recipient routes absent from the approved MATG), despite per-transfer checking. Adding the route guard eliminates this category, demonstrating the value of explicit route restriction as an integral component of TG-CI’s enforcement.

Auto-Region offers a privacy–utility trade-off through item-specific scope enforcement: it reduces leakage by 75.7% relative to Online Atomic CI with a 0.69-percentage-point utility decrease. Compared with route-guarded Online Atomic CI, it offers higher utility, alongside higher residual leakage. These results position Auto-Region as the higher-utility TG-CI variant, supporting useful processing within an admissible region.

Full-Scope provides further protection after route control: it reduces total leakage by 51.7% relative to the guarded baseline, with utility 0.69 percentage points lower. Memory-release leakage falls by 56.0%, consistent with prospective admission withholding locally admissible releases whose downstream scope is inadmissible. On-graph and final-output leakage also decreases by 42.6%.

We thank the reviewer for motivating this breakdown and will incorporate the analysis into the revised manuscript.

AuthorsOpenReview ↗

Author Response (2/2): Additional Experimental Analysis and Benchmark Validation

W4&Q3. Benchmark annotation and independence.

We thank the reviewer for highlighting a missing clarification. YY specifies the intended disclosure policy: whether a recipient may receive an information item for its assigned task. One researcher drafted and checked the labels, and a second performed verification of the proposed labels, using task requirements, common-sense privacy expectations, and relevant legal principles; therefore, we do not report an inter-annotator agreement statistic from this procedure. The criterion is legitimate information need rather than TG-CI’s decisions. Task conditioning therefore defines the privacy objective; it does not itself make the evaluation circular or prescribe a particular enforcement mechanism.

Crucially, Online Atomic CI uses the same task-conditioned atomic CI judgment at every realized memory and inter-agent transfer (§5.2). It receives the recipient’s task and transmission context, considered alongside information content, applicable norms, and disclosure history. TG-CI additionally reasons prospectively over downstream propagation and enforces the resulting scope. The evaluation compares actual unauthorized disclosures during execution against the same policy labels. We will clarify the annotation procedure and the information available to each baseline in the revision.

Q4. Results against an adaptive compromised agent

Thank you for this suggestion. We conducted an additional, preliminary stress test in which compromised agents were prompted to paraphrase protected information before forwarding it. Mean unauthorized disclosure counts were:

MethodParaphrasing OFFParaphrasing ONAbsolute increase
Auto-Region1.4881.727+0.239
Full-Scope0.6240.637+0.013

Both modes show modest degradation under this test, with a smaller increase for Full-Scope. Our conservative cumulative-CI accounting records the original items authorized to reach each recipient, independently of their subsequent wording. TG-CI's protection is not solely dependent on the semantic frontier gate, however, paraphrasing can challenge the MiniLM/NLI gate’s recognition of protected content; the observed degradation is consistent with this enforcement limitation.

This experiment uses prompt-induced paraphrasing without a controlled attack budget. Our contribution centers on the TG-CI formulation, framework, and proof-of-concept validation. Systematic evaluation under controlled adaptive-attack budgets and strengthening semantic enforcement are priorities for future work. We will include these results and clarify the evaluation scope and limitations in the revised manuscript.

Q5. Higher utility in Local-History CI under Haiku-Luna.

Thank you for raising this point. We rechecked the runs and confirmed the reported values. No Protection is not an oracle upper bound on task utility: it removes privacy filtering, but does not guarantee that agents retrieve or use the correct information. Agents may therefore retrieve distractors, omit necessary facts when forwarding, stop retrieval too early, or select an incorrect answer. Local-History CI changes the context presented to agents by suppressing some inappropriate items, which can alter subsequent retrieval and communication decisions and occasionally improve task accuracy. The observed difference is modest, and we do not interpret it as a general utility advantage of privacy filtering.

Thank you again for your valuable comments and constructive suggestions. We have carefully addressed your concerns in our rebuttal and provided additional clarifications where necessary. As the discussion period is limited, we would be most grateful if you could find a moment to look over our response and let us know whether our explanations have adequately addressed your questions. Should any concerns remain, we would be very happy to provide further clarification. Since our experiments take considerable time to complete, an early reply would be of great help in allowing us to prepare any additional results you may wish to see within the remaining discussion period. We sincerely appreciate your time and effort in reviewing our paper.

Reviewer_Sd7KOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method and benchmark for privacy failures in multi-agent LLM systems that local disclosure checks miss. The paper proposes judging each memory release over the agent's planned task graph. An item is released only with an authorized propagation scope. Every planned hop is checked by an LLM contextual-integrity judge conditioned on that recipient's disclosure history. The paper studies Two enforcement modes are studied. First, blocking at the source if any downstream transfer is inadmissible. Second, releasing within the maximal admissible region with the item stripped at the region's frontier by a local semantic filter. On a new 144-case benchmark, both modes sharply reduce unauthorized disclosures.

Critical strengths
  1. Most memory items are distractors, so indiscriminate blocking is detectable. The two-fact questions make over-removal show up as lost accuracy. Ground truth is item-level and includes annotated benign pairs for cumulative violations.
  2. The qualitative result holds across three agent–judge model assignments and three graph shapes with a stable method ordering.
Critical weaknesses
  1. Absense of a strong baseline. Every baseline is the author's own variant of the same judge and the the headline reduction is measured against their own reference point.
  2. Ground truth and measurement both use and depend on LLMs. The benchmark is also LLM-generated with unspecified researcher review. Trace scoring also uses an LLM from the same family as the primary judge.
  3. With weaker judge, the strict mode is only about 62%, which way far behind 93-96%. The trade-off clearly depends on judge strength.
Questions for authors
  1. Can we have atleast one published defense be adapted as an external baseline or maybe even a a cost-matched atomic baseline given the same token budget as the proposed modes?
  2. What happens when a compromised agent paraphrases protected content instead of forwarding it? Also, how does the method behave if the task graph changes mid-execution?
3 comments
AuthorsOpenReview ↗

Author Response

We thank the reviewer for the constructive feedback. We address the baseline, measurement, judge-dependence, and robustness concerns with additional analyses.

W1/Q1. Comparison with published defenses.

We thank the reviewer for suggesting stronger external baselines. We additionally adapted two published defenses to our multi-agent setting: FIDES-based label IFC (Costa et al., 2025) and AirGapAgent-style data minimization (Bagdasarian et al., 2024).

The FIDES-based baseline was supplied with oracle per-item reader permissions from the benchmark policy matrix, which TG-CI does not access during execution. Nevertheless, it incurred 1.688 unauthorized disclosures per run, mainly contribute to cumulative violations arising from combinations of individually permitted information. The AirGapAgent-style baseline incurred 15.652 disclosures, compared with 1.488 for TG-CI Auto-Region and 0.624 for Full-Scope, and was unable to prevent much local, cumulative and transitive leakage. (Also see detailed results in our response to KAfd - W2&Q2 and Dh2Q - W9).

These results provide external comparisons beyond our controlled baselines. We will incorporate the experimental settings and results into the revised manuscript, while clarifying that these are adaptations of the published approaches.

W2. Benchmark provenance and LLM-assisted scoring.

MAPLE-Bench uses LLM-generated synthetic cases, but the item–recipient policy labels Y were drafted and checked by one researcher and verified by a second against assigned tasks, privacy expectations, and relevant principles. The Y is independent of the runtime CI judge and is withheld from TG-CI. All methods are evaluated against the same policy matrix using an identical trace-scoring pipeline, ensuring a consistent evaluation protocol across baselines and TG-CI. We acknowledge that LLM-assisted trace matching, particularly with an overlapping model family, can introduce correlated scoring errors. We will clarify the generation/annotation/scoring pipeline and explicitly acknowledge this LLM usage limitation.

We also tested on the published MAMA (Liu et al., 2026) benchmark (150 runs): mean disclosures were 0.093 for Online Atomic CI, 0.060 for Auto-Region, and 0 for Full-Scope (see detailed results in our response to Dh2Q - W8). MAMA measures a different extraction-oriented threat. The findings provide complementary evidence of TG-CI's privacy–utility benefits, although the two benchmarks are not directly comparable due to their different evaluation objectives. We will include the results and adaptation details in the revision.

W3. Dependence on judge strength.

We agree this is a meaningful limitation. We measured Full-Scope task-necessary item withholding rates: 18.45% (Claude/GPT), 34.16% (Claude/Claude), and 17.07% (GPT/GPT). With Claude agents fixed, the Claude judge exhibits substantially higher false blocking and lower Full-Scope utility, consistent with its lower atomic-judgment recall (78.87% versus 93.30%). The stronger whole-scope admission can also amplify errors from a weaker judge. The findings suggest that judge reliability is an important factor in the privacy–utility trade-off, motivating improvements in judge capability and alignment. We will emphasize judge calibration and the resulting privacy–utility sensitivity.

Q2. Paraphrasing and changing graphs.

In a preliminary prompt-induced paraphrasing test, Auto-Region leakage rose from 1.488 to 1.727, and Full-Scope from 0.624 to 0.637. While both variants retain substantial containment, we acknowledge fixed MiniLM/NLI thresholds remain vulnerable to semantic evasion, and this preliminary test does not establish robustness against fully adaptive adversaries. Full-Scope is less reliant on that filter because it withholds whole-scope-inadmissible releases at the source.

Our current implementation assumes a predefined, trusted MATG and messages on routes absent from that approved graph are blocked. Supporting dynamic replanning or graph updates is left to future work as stated in the paper. It would require an additional trusted graph-monitoring mechanism to detect topology changes, update propagation authorizations, and maintain consistent disclosure states. The focus of this paper is TG-CI's formalization, framework, and proof-of-concept validation, while dynamic authorization updates and stronger adversarial robustness are for future work. We will clarify this assumption and future direction.

Thank you again for your valuable comments and constructive suggestions. We have carefully addressed your concerns in our rebuttal and provided additional clarifications where necessary. As the discussion period is limited, we would be most grateful if you could find a moment to look over our response and let us know whether our explanations have adequately addressed your questions. Should any concerns remain, we would be very happy to provide further clarification. Since our experiments take considerable time to complete, an early reply would be of great help in allowing us to prepare any additional results you may wish to see within the remaining discussion period. We sincerely appreciate your time and effort in reviewing our paper.

Thank you authors for the detailed comment. The new external baselines and MAMA results address my concerns with baseline partly. The withholding rate analysis is a good explanation of the judge dependence. The same family scoring concern remains, but is acknowledged. I would like to retain my previous scores for the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

We thank all reviewers for their constructive feedback. We have uploaded a revised manuscript, expanding the main text from 9 to 10 pages. Substantive additions and modifications are highlighted in blue for ease of review. Below, we summarize the changes corresponding to the reviewers' concerns.

1. Novelty and relationship to prior work.

  • §1 (pp. 2–3): Clarified TG-CI's contribution and its distinction from traditional IFC.

2. Formal definitions and assumptions.

  • §3 (pp. 4): Explicitly defined the disclosure-history space and accumulation operator.

3. External baselines.

  • §5.7.1 (p. 10): Added adaptations of two published defenses: FIDES-based label IFC and AirGapAgent-style minimization.

4. Atomic CI comparison.

  • §5.3 (pp. 7–8), Appendix F (pp. 18–19): Added a delivery-channel decomposition and route-guard ablation.

5. Benchmark construction and evaluation reliability.

  • §5.1 (pp. 6–7): Clarified that benchmark policy labels were drafted and checked by one researcher and verified by a second, independently of the runtime CI judge.
  • §5.2 (p. 7): Explicitly identified LLM-assisted trace matching and clarified that all methods use the same scoring pipeline.

6. External benchmark validation.

  • §5.7.2 (p. 10): Added evaluation on 150 MAMA benchmark cases.

7. Judge reliability and privacy–utility sensitivity.

  • §5.5 (p. 9): Added task-necessary item withholding rates.

8. Threat-model transparency, paraphrasing, and dynamic graphs.

  • §2 (p. 3), Appendix E: Clarified the adversary model and documented the compromised-agent attack configuration.
  • Appendix G (p. 19): Added a prompt-induced paraphrasing stress test.
  • §4 (p. 5), §6 (p. 10): Made the predefined, trusted MATG assumption explicit and identified dynamic authorization updates and stronger adaptive-attack robustness as future work.

We hope these revisions address the reviewers' principal concerns. We sincerely appreciate the reviewers' feedback and suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.