acceptodds
Under review as a conference paper at ICLR 2027

Policy-State Fragmentation in Multi-Agent LLM Systems: Analyzing Binding Continuity and Runtime Enforcement

Abstract

Useful privacy enforcement requires policy state that distinguishes permitted from forbidden actions at the point of execution. We identify and formalize context-fragmented violations, distinguish them from semantic source-binding loss, and contribute layered diagnostics and Distributed Sentinel, a reference enforcement framework. The central finding is that policy-state availability supports utility as well as safety. Independent annotators are blind to predictions; supplying policy state mainly resolves abstentions in a post-label diagnostic, raising decision coverage from 65.8% to 98.3%. In controlled certified pairs, supplied bindings enable 64/64 permitted sends and prevent 64/64 forbidden sends across matched mechanisms. In a three-way plan diagnostic with neutral endpoint names, GPT-5.4 approves 37/160 labeled-risky plans locally and 0/160 with full policy. A separately frozen structured-handoff diagnostic retains every binding through 72 transformations, yet actors propose seven prohibited sends, all intercepted. Binding preservation and action compliance are thus distinct. Separately, under a forced-binary contract, augmentation with directives and registry facts increases safe-task refusals for four of five models. Fixed-output replay isolates interface-induced stops in an external mapping pipeline. These diagnostics separate information availability, binding continuity, and pre-effect placement. We recommend evaluating safe-task completion alongside violation prevention.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multi-agent LLM systems, AI agent security, privacy policy enforcement, data provenance, information-flow control, runtime enforcement

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_8riiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies privacy failures caused by policy information being split across agents. It defines context-fragmented violations and distinguishes them from losing source-policy bindings during a transformation. Distributed Sentinel carries identified source bindings and checks actions before execution. The experiments separately examine missing information, binding preservation, enforcement placement, and mapping errors. Supplying policy information raises decision coverage from 65.8% to 98.3% on 120 independently annotated cases. On 64 controlled allowed/forbidden pairs, Sentinel and matched controls make all decisions correctly. A separate 126-case study finds substantial false blocking in the frozen mapping pipeline; renaming fields in saved outputs explains many of these errors.

Critical strengths

S1. The paper distinguishes missing policy information from losing a source binding during transformation. Definition 2.2 also keeps the policy check fixed between local and global views, so the mismatch cannot be explained simply by applying different rules.

S2. The Q1 analysis separates abstention from an incorrect decision. Of the 41 newly correct answers, 38 were previously abstentions. This gives a more precise account of what supplying policy information actually improves.

S3. The external mapping study identifies an interface problem by renaming keys while keeping the saved model outputs otherwise unchanged. This removes 40 false blocks and introduces one slip. The authors retain the poor original result and clearly separate the later repair from independent validation.

Critical weaknesses

W1. The conceptual advance over existing information-flow control is limited. Context-fragmented violations describe decisions made with incomplete policy information, while source-binding propagation and pre-effect checks follow established enforcement principles. The separation of failure types is useful, but I am less convinced that it amounts to a substantial new security model.

W2. The framework preserves bindings that have already been identified correctly. It does not determine which sources still govern a rewritten artifact or when a restriction can be released. Conservative source union can therefore retain irrelevant restrictions, leaving a gap between authenticating policy metadata and deciding whether that policy applies to the content.

Questions for authors

What security behavior is specific to the community setting beyond a persistent assistant applying contextual access-control rules across sessions?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 8rii

Thank you for the careful reading, particularly the distinction you draw between preserving a binding and deciding whether it still applies after rewriting.

W1: What conceptual contribution goes beyond incomplete-information decisions and established IFC enforcement?

We distinguish failures that call for different changes to an LLM workflow. For example, 53/54 false blocks in the external pipeline are missing-input exits; 56/64 reference-BLOCK cases stop there too. Renaming keys in the saved outputs removes 40 false blocks and introduces one slip, without asking the model for another answer. This locates a repair at the input interface.

Definition 2.2 fixes the predicate, action arguments, executed prefix, missing-input convention, and execution boundary. Only the available context changes. This lets us attribute a discrepancy to missing information without also changing the rule or the execution history. The 64 certified allowed/forbidden pairs provide executable examples: each pair has identical downstream observations but opposite full-state permissions. Registry, CaMeL, and Sentinel agree when given the full state under the tested predicates.

The definitions also separate two testable conditions. A CFV requires every action to pass its local check and at least one to violate the full-state check under the same contract. Semantic laundering requires retained source meaning with a lost binding. Neither implies the other: policy state may never have been available locally, and a lost binding may concern an ultimately permitted use. This determines which observations an experiment must collect.

The experiments connect these distinctions to specific interventions:

Failure locationEvidenceCorresponding intervention
Binding loss in free-form handoffsTwo readers identify retained meaning with unavailable bindings in 87/144 source items across 26/36 outputsCarry identified bindings through handoffs
Action choice with intact bindingsAll bindings survive 72 structured transformations; seven prohibited sends are proposed and all interceptedCheck proposed actions before execution
Mapping interfaceOn 56 fresh cases, the frozen repair reduces allowed-task stops from 13/28 to 1/28; complete-workflow exact matching reaches 51/56Repair the mapper and validate on fresh workflows

The free-form handoffs forward preceding outputs in full, without host code stripping metadata. Raw reader agreement is 84.7% for meaning and 98.6% for bindings. We retain agreed labels and mark disputed fields UNCERTAIN, keeping all items. These are controlled tasks: 144 source items are nested within 36 outputs and 18 workflow pairs. The fresh-case results use the scenario authors' construction references, with 0/28 prohibited-task approvals before and after repair and the same downstream enforcement engine.

Sentinel records bindings, decisions, and effects to check each diagnosis and its corresponding intervention.

W2: How are semantic applicability and restriction release handled, given conservative source union?

In our notation, Dep(d) denotes semantic dependencies and T(d) the recorded bindings. The preservation result concerns T(d). End-to-end enforcement also requires complete relevant bindings, faithful policy encoding, and trusted interception; authentication establishes metadata integrity.

The new readers check whether source-attributable facts survive and whether the output still links them to the relevant recipient restrictions. The runtime retains recorded bindings conservatively, which can retain a restriction even when its source is no longer semantically relevant. Releasing it would require a separately validated semantic declassification rule. We therefore report permitted completion alongside blocking. In the paired supplied-binding study, 64/64 permitted sends complete and 0/64 forbidden sends are delivered.

Q1: What security behavior is specific to a community beyond a persistent assistant across sessions?

The relevant features are multiple policy owners, different local views, and handoffs between them. In the paper's Titan example, R&D owns the disclosure restriction, while marketing receives a derived summary. The publication check still needs that restriction even though marketing cannot access the source record. A persistent assistant with the same separation of state faces the same issue. If that assistant retains the relevant policy state at every checked action, this particular local/full-state mismatch is absent. The community setting makes the separation concrete: one participant owns the policy, another transforms the content, and a third may execute the action. The criterion requires applicable source-owned policies to be available at that final boundary, regardless of how many model instances implement these roles.

Thanks again for your review.

Thank you for the response! I agree with you that separating missing policy info from lost bindings can help to identify problems. But, the concern about conceptual contribution still the same. The distincition truely useful for diagnosis. However, I still think that the paper adds little to existing info flow control methods. If you give me more evidence, I'm happy to raise the point.

Thank you for clarifying your concern.

Our methodological contribution is to locate policy-enforcement failures through controlled comparisons of available information, recorded bindings, and input interfaces. Definition 2.2 specifies what must remain fixed when comparing local and full-state checks.

The interface replay asks a separate question: does the rejection persist when the same model response is supplied through corrected fields? The fresh-case study then tests whether a mapper repair improves task completion. The results reported in our rebuttal are summarized below.

StepEvidenceWhat it supports
Locate the failure in the original 126-case studyMissing inputs account for 53/54 false blocks and 56/64 reference-BLOCK cases.The same failure produces incorrect rejection and apparently correct blocking; the final verdict alone does not establish successful policy reasoning.
Test the diagnosis on saved outputsRenaming fields removes 40 false blocks and introduces one slip, without changing model responses.The model responses stay fixed, so the corrections identify a problem at the input interface.
Evaluate the frozen schema-aware mapper on 56 fresh casesAllowed-task stops: 13/28 → 1/28; strict permitted completion: 0/28 → 25/28; prohibited approvals: 0/28 → 0/28; full-workflow exact match: 0/56 → 51/56.The repair improves permitted completion on fresh cases while the downstream enforcement stays unchanged.

The last row compares the original and repaired mapper prompts using the same model snapshot and downstream engine. It uses the scenario authors' construction references, separately from the original independent study. Strict completion requires an ALLOW decision and an exact match to the full specified workflow. Fixed-output renaming and the fresh-case mapper evaluation are distinct interventions.

The structured-handoff study examines a different stage: all bindings survive 72 transformations, yet seven prohibited sends are proposed and intercepted. This study tests action checking with intact bindings; the interface study tests whether the checker receives the required inputs.

Thank you again for taking the time to discuss these results with us.

Thank you for the response. I am more clearly on how your method helps find where a policy check fails. I still think the paper adds little to existing work on information flow control. I will remain unchange.

Reviewer_d31GOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors introduce the concept of context-fragmented violations and semantic laundering under multi-agent settings. The core assumption is that individual agents pass local security policy checks because policy bindings attached to source data are lost after downstream rephrase. To avoid losing the binding, the authors propose DISTRIBUTED SENTINEL using taint tokens that propagate the union of source bindings. Then, the authors run evaluation to access whether model can provide correct judgement with source policy and whether DISTRIBUTED SENTINEL can be helpful in external cases.

Critical strengths

S1. It is good to report the negative results for the proposed method and reveal the protocol sensitivity.

Critical weaknesses

W1. No empirical observation of natural policy loss. The authors never observe or measure whether agents naturally lose policy labels in real multi-agent workflows. In fact, all cases tested are artificially constructed to strip policy bindings. Given these oriented cases, restoring missing policy facts improves decisions is true by definition.

W2. There is no advantage over baselines. DISTRIBUTED SENTINEL provides taint tracking and shows no performance advantage over basic baselines.

W3. On the only independent external evaluation, the E2E pipeline achieves only 57.1% accuracy with an 87.1% false-block rate.

Questions for authors

Q1. Given that DISTRIBUTED SENTINEL ties the simpler Registry checker and CaMeL baseline, what is the specific technical advantage of the proposed method?

Q2. Can the authors provide any empirical evidence of semantic laundering or policy label loss occurring in an non-controlled multi-agent system?

3 comments
AuthorsOpenReview ↗

Response to Reviewer d31G

Thank you for looking closely at the assumptions and external results. We have added a handoff study and tested the frozen repair on fresh cases.

W1: Are the cases constructed to strip bindings, making the benefit of restoring policy information inevitable?

The submitted structured-handoff study preserves all bindings through 72 transformations, yet actors propose seven prohibited sends, all intercepted. The evaluation therefore includes workflows with intact bindings.

Section 2 also distinguishes two situations. A CFV can occur because relevant policy facts were never available in the acting agent's local view. Semantic laundering requires source-derived meaning to survive while its binding is lost. Binding loss is not a premise of every experiment; Q2 reports the new measurements of loss during free-form rewriting.

W2: Why does Sentinel match simpler Registry and CaMeL controls rather than outperform them?

In the matched study, Registry, CaMeL, and Sentinel make the same decisions with the same supplied policy state. The comparison fixes the policy rule so we can examine the role of information availability. In the 64 certified pairs, the downstream observations are identical within each pair, but source permissions differ. Supplying the full state lets all three mechanisms complete 64/64 permitted sends and prevent 64/64 forbidden sends.

On the same pairs, the permissive local control executes 64/64 forbidden sends, while the fail-closed local control completes 0/64 permitted sends. These constructed pairs expose the safety–utility tradeoff when the local observation cannot distinguish the two permissions; supplying the missing state resolves both outcomes under the same policy rule.

W3: How should we interpret 57.1% external accuracy and an 87.1% false-block rate?

On 56 fresh cases, the frozen repair reduces allowed-task stops from 13/28 to 1/28 (46.4% to 3.6%), with 0/28 prohibited-task approvals in either condition. The model and downstream engine are unchanged, both prompts were fixed before evaluation, and all cases are included. Against the scenario authors' construction targets, verdict agreement improves from 43/56 to 55/56. Strict permitted-workflow completion rises from 0/28 to 25/28; this requires an ALLOW decision and an exact match to the full author-specified workflow JSON. Complete-workflow matching improves from 0/56 to 51/56.

Both contributors' cases improve: verdict agreement rises from 19/28 to 27/28 for A and from 24/28 to 28/28 for B. The 12 corrections cover six policy families, with no verdict regressions; the seventh family remains 8/8. These author-reference results are reported separately from the original 126-case independent study.

The original result remains 57.1% operational correctness and an 87.1% false-block rate. Of its 54 false blocks, 53 stop on missing context; 56/64 reference-BLOCK cases stop there too. Renaming fields in the saved outputs removes 40 false blocks and introduces one slip, without changing the model outputs. Because the outputs stay fixed, these corrections identify an input-interface problem. The fresh-case test shows that the frozen repair also improves permitted completion with the downstream enforcement rule held fixed.

Q1: What is the specific technical advantage over Registry and CaMeL?

For the tested predicates, the result is equivalent enforcement. Our contribution is the same-check definition and experiments that locate failures across the workflow, supported by Sentinel's binding transport and pre-effect checks. Sentinel authenticates the recorded metadata and retains decision and execution traces. This lets us distinguish a missing binding from a prohibited proposal with intact bindings or a mapper error, and test the corresponding repair. The matched accuracy is the control for this analysis, rather than evidence of a stronger decision rule.

Q2: Is there evidence of semantic laundering or policy-label loss in a non-controlled multi-agent system?

Our new measurements use controlled tasks with free-form model handoffs. We collected new summaries and rewrites on all 36 existing tasks. Each downstream actor received the preceding output verbatim, including metadata; no prompt requested label removal. Two independent readers found retained meaning with unavailable bindings in 87/144 source items across 26/36 outputs. They saw the sources and outputs, but not system decisions. Raw agreement was 84.7% for meaning and 98.6% for bindings. We retained every item, kept agreed labels, and marked disagreements UNCERTAIN. Items are nested within 36 outputs and 18 workflow pairs. These are observed examples of loss during model rewriting without deliberate label stripping; they are not observations from a non-controlled deployment.

Thanks again for your review.

Thank you for the detailed rebuttal and additional experiments! I appreciate the efforts made by authors to address my concerns. The response partly resolved my concerns. The left main concerns are still "why I should use your technique" (advantage) and "when I should use it" (practical). I am looking forward to know more details about the new measurements consists of controlled tasks with free-form model handoffs.

Since the main concerns still exist, I would like to keep my original score.

Thank you for the follow-up. We would like to address the intended use and give more detail on the free-form handoff study.

When would this be useful?

The intended setting is a workflow where one agent reads source records, another rewrites the content, and a downstream agent proposes an external action without access to all source-owned restrictions. For example, a summary may retain a project fact while no longer identifying who may receive it. The execution check still needs that restriction.

What does Sentinel provide?

Trusted instrumentation carries identified source bindings across transformations, and the interceptor checks the proposed action against the applicable policies before execution. The source restrictions therefore remain available to the checker even when the generated prose no longer repeats them. This avoids relying on the downstream agent to reconstruct those restrictions from the summary.

How were the free-form handoffs measured?

We generated new outputs on all 36 existing controlled tasks, arranged in 18 workflow pairs. An analyst received source records with recipient restrictions and was asked for a brief of about 100 words. An editor was asked to rewrite the complete brief in about 80 words; a dispatcher then proposed an action. Both writing prompts asked the actors to preserve facts and respect handling restrictions.

Every response was forwarded in full, including any metadata. No prompt requested label removal, and the host code did not strip labels.

Two independent readers saw the sources and final edited outputs, without the action proposals or checker decisions. They checked whether source-attributable facts survived and whether the output linked those facts to the source's recipient permissions. Bindings could appear in prose or metadata; no particular output format was required:

MeasurementResult
Source–artifact items evaluated144
Retained meaning with unavailable binding, agreed by both readers87/144
Outputs containing at least one such item26/36
Raw agreement on meaning retention122/144 (84.7%)
Raw agreement on binding availability142/144 (98.6%)

We retained all items and used a post-hoc field-wise agreement rule: agreed labels were kept; disagreements were marked UNCERTAIN. The items are nested within the 36 outputs and 18 pairs.

For example, in dyn_032 the edited brief retains “Package 3-4-4 has 46 checks completed” but omits that source's recipient restriction. The dispatcher proposes sending to partner_digest, whereas the source permits only approved_internal_partner. The example shows exactly which fact survives and which authorization information is missing at the action boundary.

These are controlled observations of model rewriting, rather than a deployment sample. They identify a concrete use case for binding transport: the task content reaches the next agent, but the associated source restriction does not.

Thank you again for taking the time to discuss these questions with us.

Reviewer_c4LLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies policy-state fragmentation in multi-agent LLM systems, distinguishing missing policy information, loss of source–policy bindings, and action compliance. It presents Distributed Sentinel, a reference framework for binding propagation and pre-effect enforcement, as well as controlled execution studies.

Critical strengths
  • I agree with their problem statement, and also appreciate their careful failure attribution. The evaluation distinguishes policy-based rejection from abstention, missing inputs, and parser failures, which is quite missing in the field. This prevents aggregate blocking rates from being mistaken for successful policy reasoning.

  • Their controled experiments are informative and valuable. Fixed-output field renaming removes 40 of 54 false blocks without changing model responses, isolating an interface failure. Blinded annotation and matched enforcement controls further strengthen the analysis.

Critical weaknesses
  • Limited incremental insight beyond established IFC principles: source-binding propagation and pre-effect checks largely instantiate existing mechanisms, and matched Registry or CaMeL controls reproduce the tested enforcement outcomes. The diagnostic methodology is useful, but I think the resulting conclusions remain mostly expected consequences of information availability and mediation placement.

  • The practical failure-to-remedy guide is still missing. No study independently establishes retained semantic dependence, lost bindings, and a resulting violation prevented by binding restoration (all structured-handoff transformations instead preserve their bindings). The external mapping study offers limited complementary validation, but its successful repairs are evaluated only on previously exposed cases.

Questions for authors
  • Could you please provide independently checked cases where source-derived meaning survives a transformation, its binding is lost, and restoring that binding prevents an otherwise unsafe action?

  • with the repaired schema and mapper frozen, would the improvements generalize to fresh workflows under matched enforcement conditions? (measured by unsafe effects and successful completion of permitted tasks)

  • Can you separate the effects of adding policy facts vs general directives to explain when additional policy context improves utility and when it increases false blocking?

1 comment
AuthorsOpenReview ↗

Response to Reviewer c4LL

Thank you for the careful review. On 56 fresh cases, the frozen repair reduces allowed-task stops from 13/28 to 1/28. Its 12 verdict corrections cover both contributors and six policy families, with no regressions against author references.

W1: What insight goes beyond established IFC and the expected effects of information availability?

Our same-check definition separates information differences from changes in the policy rule. The experiments distinguish binding loss, prohibited proposals with intact bindings, and mapping failures. Registry/CaMeL provide matched full-state controls; the studies below test the corresponding repairs.

W2: What is the failure-to-remedy guide, and does validation extend beyond exposed cases?

Q1 reports the binding study; Q2 tests the repair on fresh cases:

Observed failureTargeted intervention
Meaning survives, binding unavailableCarry identified bindings through trusted handoffs
Bindings intact, prohibited action proposedCheck the action before execution
Missing-input stops corrected by key-only replayRepair the mapping interface and test on fresh cases

Q1: Can independent checks show retained meaning, lost bindings, and an unsafe action prevented by restoration?

Two independent readers identified retained meaning with unavailable bindings in 87 of 144 source–artifact items, covering 26/36 outputs. We generated new outputs on all 36 existing controlled tasks, with two free-form transformations followed by a dispatcher. Each actor received the previous output in full, including any metadata. We did not instruct the actors to remove labels.

Readers saw the sources and outputs, but not proposals or checker decisions. Their raw agreement was 122/144 (84.7%) for meaning and 142/144 (98.6%) for bindings. We used a post-hoc agreement rule: keep agreed labels and mark disputed fields UNCERTAIN. All 144 items remain included; the 36 outputs form 18 workflow pairs.

For example, dyn_032 retains “Package 3-4-4 has 46 checks completed” but omits its recipient restriction. The dispatcher proposes sending to partner_digest, while that source permits only approved_internal_partner. Adding the confirmed missing bindings changes the guard from INSUFFICIENT_INFORMATION to BLOCK, with the text and proposal held fixed. Two of the three eligible external-send proposals change this way; one remains insufficient. Identical checker requests reuse saved first responses. No prohibited relay delivery occurred, so a complete unsafe-execution/restoration chain was not observed.

Q2: Does the frozen schema-aware mapper generalize to fresh workflows in safety and permitted-task completion?

On 56 fresh cases, allowed-task stops fall from 13/28 to 1/28 (46.4% to 3.6%), with 0/28 prohibited-task approvals in both conditions. We compared the original and already frozen schema-aware prompts using the same model snapshot, policy descriptions, and downstream engine. All 56 cases and 112 first responses are included. The references are the two scenario authors' construction targets and specified workflows.

MetricOriginal mapperSchema-aware mapper
Author-target verdict agreement /564355
Allowed-task stops /28131
Prohibited-task approvals /2800
Allowed-workflow strict completion /28025
Complete-workflow exact match /56051
Missing-input stops /56321

Strict completion requires an allowed target, an ALLOW decision, and an exact match to the full author-specified workflow JSON. The seventh policy family already had 8/8 verdict agreement and remains unchanged. Encoded unsafe effects are zero in both conditions, measured in isolated local sinks.

Q3: How do policy facts and general directives separately affect utility and false blocking?

We ran a 2×2 comparison on the same 200 complete plans, keeping the GPT-5.4 snapshot, binary decision contract, and base role policies fixed. The table uses the submission's generator labels.

Added contextSafe STOP /40Risky PROCEED /160
None81
Directives81
Registry facts125
Both131

Directives alone change one safe plan in each direction, leaving the total at 8 STOPs. Facts alone change four safe STOPs to PROCEED and eight PROCEEDs to STOP. With facts present, adding directives reduces risky approvals from 5 to 1, while safe STOPs rise from 12 to 13. Thus the two additions have different safety and utility effects in this comparison.

“Facts only” means no added global directives; the base prompts still contain role policies. The facts cover sensitivity, ownership, domain, and authorization. The table includes uniform recovery of three inline terminal decision markers and one identical-request retry after a gateway failure. We retain the strict first-attempt records.

Thanks again for your review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.