acceptodds
Under review as a conference paper at ICLR 2027

GuardCircuits: Local Composition and Global Placement for Agentic Privacy

Abstract

As multi-agent systems take on real workflows spanning agents, tools, memories, and external services, protecting sensitive information becomes increasingly important while preserving the information needed to complete the task. Many guard mechanisms have been introduced to preserve privacy. Existing work has largely focused on improving individual guards, while placement is treated locally, for example at a particular tool call or system boundary, without considering the global graph structure and the joint tradeoff between privacy, utility, latency, and cost in their design. We therefore ask a more architectural question: *what behaviors and design principles emerge when imperfect guards are composed, and where should the resulting compositions be placed in an agent workflow?* We introduce *GuardCircuits*, a framework inspired by electrical circuits, that composes guards using series and parallel structures together with deterministic or stochastic switches. We show that a guard circuit admits an equivalent privacy–utility–latency–cost signature and, in a high-reliability regime, a tropical (min–plus) resistance law that reveals depth–width tradeoffs between privacy and utility. We then study global placement: ideal mediation reduces to a directed cut problem, while imperfect guards can be arranged in ordered barriers whose privacy resistance accumulates across successive lines of defense. Empirical results on privacy benchmarks support these composition laws and show that minimum-cut placement reduces leakage, giving a principled basis for both arranging and placing guards in agentic systems.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, multi-agent systems, guardrails, privacy, AI safety, Information leakage

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_GHP8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how imperfect privacy guards should be composed and placed in multi-agent workflows. The authors introduce GuardCircuits, a framework that models guards as components with privacy, utility, latency, and cost signatures. They derive composition laws for serial, parallel, and switched guard circuits, including a high-reliability min-plus law that explains depth-width tradeoffs. They also study global placement, showing that ideal guard placement reduces to a directed minimum cut and that stochastic guards can form ordered barriers across repeated executions. Experiments on three privacy benchmarks evaluate serial vs. parallel guard composition and min-cut vs. random guard placement.

Critical strengths
  1. Privacy guards are often imperfect, and their placement in agent workflows can matter as much as the guard model itself. The research question is interesting.
  2. The theoretical framework is clean and potentially useful.
  3. The distinction between local guard composition and global workflow placement is valuable.
Critical weaknesses
  1. The empirical validation is relatively limited compared with the scope of the theoretical claims. The experiments mostly confirm intuitive monotonic trends, and does not convincingly show that the proposed signatures can be accurately calibrated and then used to predict performance of unseen guard circuits or new workflows.
  2. The placement experiments use constructed workflows with known two-edge minimum cuts and relatively simple topology. This supports the proof-of-concept, but it is unclear whether the method improves over practical baselines in real agent systems where the communication graph, information origins, and tool side effects may be only partially observed.
  3. The evaluation measures leakage only on released outputs in several experiments, while the motivation is system-level privacy in multi-agent workflows. This creates a gap between the stated concern about internal paths, memories, and tool channels and the empirical metric used to validate the framework.
Questions for authors

See the Weaknesses.

4 comments
AuthorsOpenReview ↗

Response 1/2

Thank you for your thoughtful feedback on our paper. We are encouraged that you found the research question interesting and the theoretical framework clean and potentially useful, and that you appreciate the distinction between local guard composition and global workflow placement. Below we have provided responses to your questions, including new experiments that compare the circuit-law predictions with observed results. Please let us know if you have any follow-up questions.


Q1. Do calibrated signatures predict unseen circuits and workflows, rather than only monotone trends?

Thank you for this suggestion. Based on your feedback, we have added two experiments that test exactly this: we calibrate the signature of a single primitive guard, and then use the circuit law (Theorem 4) to predict the behavior of circuits that were never used for calibration (App. F.8 in the revised manuscript). In both experiments, the signatures are estimated only from single-guard (k=1k=1) calls on each case, using the GPT-5-mini censoring guard.

Pure series and pure parallel circuits (App. F.8.1, Figure 17). We predict leakage, task success, latency, and token cost for series and parallel circuits with k=2,…,5k=2,\dots,5 guards on all three datasets. The predicted curves closely track the observed curves in most settings. The main exception is the pure series circuit on ACIB, where observed latency and token cost fall below the prediction: the first guard removes much of the input's complexity, so later guards process shorter messages and are faster and cheaper than the k=1k=1 calls used for calibration.

Unseen mixed circuits (App. F.8.2, Figures 18–20). We then consider circuits that interleave series and parallel blocks under a budget of six guards. We randomly sample 18 of the 64 distinct circuits and run them on all 117 ACIB cases. Using only the k=1k=1 calibration, the circuit law predicts both the ordering and the level of the observed values:

MetricMean Absolute ErrorPearson rr
Leakage0.0090.966
Task Success (Utility)0.0250.908
Latency6.03 s0.932

As in the pure series case, observed latency is somewhat lower than predicted for the slowest circuits, likely for the same reason. Overall, these results show that calibrated signatures predict the performance of unseen circuits quantitatively, not only the monotone trends.

We note that these experiments test circuit composition at a fixed placement. Predicting end-to-end leakage for new workflows additionally depends on the placement assumptions discussed in Q2 below.


Q2. Are simple constructed two-edge-cut workflows too limited for real systems?

Thank you for raising this point. We agree that the Early-Cut (EC) and Late-Cut (LC) topologies in Experiment 2 are small and constructed. We designed them this way on purpose, to isolate the placement principle. Each topology has a unique two-edge minimum cut, and the cut sits at a different stage in each one: between the propose and review stages in EC, and between the review and resolve stages in LC (Figure 8, App. E.4). The minimum cut outperforms random placement on both topologies and for every guard model (Table 3), so the effect does not depend on one particular structure.

We would also like to clarify that the placement principle does not depend on the size of the graph. A minimum cut only contains edges that lie on some path from a protected origin in OqO_q to an unauthorized recipient in UqU_q: removing any other edge from a cut leaves a valid cut with lower cost. In a large agentic system, Theorem 18 therefore only needs the subgraph of components that are reachable from OqO_q and can reach UqU_q, and EC and LC can be read as examples of such a subgraph. Computing the cut is a standard max-flow problem that runs in polynomial time [7], so graph size is not the bottleneck.

We agree that the harder issue is the one the reviewer identifies: Theorem 18 assumes that every channel between OqO_q and UqU_q is represented in GG. If a tool side effect, shared memory, or log is missing from the design-time graph, the computed cut may leave a path unguarded. We currently state this assumption under "Structural channel completeness" in App. G, and we will expand that paragraph to address partially observed workflows directly. One conservative fallback is to guard the edges entering UqU_q. This remains a valid cut as long as those edges are known, even if the internal structure of the workflow is only partially observed, but it can cost more than the true minimum cut (in EC and LC it requires 3 guards rather than 2). Designing placements that remain robust under partial observability is an interesting direction for future work.

AuthorsOpenReview ↗

Response 2/2

Q3. Why measure only released outputs when the motivation includes internal leakage?

Thank you for raising this point. Our motivation is indeed to prevent leakage from protected origins OqO_q to unauthorized recipients UqU_q, and both can be internal components of the workflow. We would like to clarify that, given the privacy policy, an unauthorized internal node is no different from an external user. Leakage is defined by the policy label χq\chi_q (Section 3), not by where a component sits in the graph: any agent, tool, or memory vv with χq(v)=−1\chi_q(v)=-1 belongs to UqU_q, and Theorem 18 cuts every path from OqO_q to all of UqU_q. For example, a shared memory that must not store medical information is protected in exactly the same way as an external user. In both cases, we measure leakage on the messages that enter the unauthorized recipient.

The external benchmarks used in the experiments were designed to measure leakage with respect to the final released output, which is why that is the main focus for the evaluation. We will make this connection between the motivation and experiments more explicit in the revised manuscript.


References

[7] L. R. Ford Jr. and D. R. Fulkerson. Maximal Flow Through a Network. Canadian Journal of Mathematics, 8:399–404, 1956. https://doi.org/10.4153/CJM-1956-045-5

Thank you very much for the response; it has addressed my concerns. I will maintain my original positive score.

Thank you for your thoughtful question and valuable suggestion, which have helped us improve the quality of our paper.

Reviewer_P1jwOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a circuit-analog study of (redaction-based) privacy guard composition in multi-agent systems, considering privacy protection, utility preservation, latency, and cost as variables of interest while deciding where and how to place such (imperfect) privacy guards, comparing combination strategies (viz. sequential, parallel), providing guidance on where and how to compose and place such guards. They provide a theoretical treatment of this problem followed by empirical evaluation over three benchmarks and multiple proprietary base AI models.

Critical strengths
  • [S1] The paper's framing of guard composition as a circuit design problem, leading up to (claimed) results viz. how a min-cut configuration can roughly match protecting all edges is intuitive and interesting. The paper attempts to answer a privacy guard circuit design question of how to position guards under a fixed guard budget.

This is interesting for composing redaction-based guards in such systems as a design problem.

Critical weaknesses
  • [W1] A major flaw is given by the authors' parallel-any composition of the outputs of guards employed in parallel: which releases a piece of information if it is allowed by any of the parallel guards (mentioned in multiple places, including in plain language in lines 408-409: "In parallel, k guards process the same input and all of their outputs are released, so a leak from any branch counts." and other discussion of parallel-any and exposed-any in their formulation. This runs counter to basic best practices: OASIS XACML 3.0, a well-known international standard for access control that defines that if any applicable policy returns deny, then that decision takes precedence over permit. This also runs counter to text anonymizers like Microsoft's Presidio, where redacted/detected spans are preserved. Indeed, it is very simply unfathomable that a privacy practitioner would adopt such a wasteful design that is leaky and allows redactions by a guard to be trivially invalidated by another guard.

    • As such, this renders the applicability and utility of results (theoretical, such as composition laws, and otherwise, such as their claimed results of how parallel composition improves utility while hampering privacy protection; such results form a large portion of those presented in this work) relying on this flawed notion of parallelism invalid for realistic privacy deployments. As such, this is a major flaw in the paper that invalidates much of its discussion and results.
  • [W2] Multiple soundness, presentation, and implementation issues that appear to be unfit for presentation at a top conference in its current state.

    • The redaction guards used in their evaluation appear to be ad hoc and not compared against existing baseline guards. At the very least, their empirical results should clearly adopt/describe how they adopt the state-of-the-art guards for this setting, else this study may rely on flawed implementations, which undermines trust in their empirical claims.
    • The authors' phrasing and framing appear to claim that they provide a framework for all privacy guardrails. A closer reading reveals that the only guards they consider are redaction-based guards. The authors need to be clear and explicitly scope their claims to this.
    • Much of the notation like FJψ,ϕFJ_{\psi,\phi} is defined in an arcane, cascading manner that is hard to follow for a reader. DW(k,m;g)DW(k,m;g), FϕJ,ηd,TˉdF^J_\phi, \eta_d,\bar T_d are defined but not used in discussion/analysis. Additionally, the authors provide theorems without a plain text discussion of what it means and how it is of interest to a privacy researcher/reader in general. The presentation in this work is genuinely hard to follow and not for the lack of careful reading (which I have done a couple of times very thoroughly) and is not of a standard that a top conference paper should aspire to achieve. It would be beneficial to clear up redundant single-use notation, proofread the paper for readability, and explain the premise and utility of a theorem carefully. Terms like primitive calls, writer-reviewer loop, typed two-terminal guard circuits (what are the types?), primitive guards, pareto synthesis, etc. are not defined or underdefined. If defined in the appendix while being used in the main text, the authors do not mention where it is defined, and it is not good practice to use terms prior to their definition being provided.
Questions for authors

Can the authors address my concerns clearly and exhaustively? These are all clarifications and/or corrections that are necessary for helping me reassess the paper, and are inclusive of the critical flaws mentioned above. This ranges from providing clarifications to improving presentation and technical flaws of the work.

  • [W1] The authors' phrasing and framing appear to claim that they provide a framework for all privacy guardrails. A closer reading reveals that the only guards they consider are redaction-based guards. The authors need to be clear and explicitly scope their claims to this.
  • [W2] A major flaw is given by the authors' parallel-any composition of the outputs of guards employed in parallel: which releases a piece of information if it is allowed by any of the parallel guards (mentioned in multiple places, including in plain language in lines 408-409: "In parallel, k guards process the same input and all of their outputs are released, so a leak from any branch counts." and other discussion of parallel-any and exposed-any in their formulation. This runs counter to basic best practices: OASIS XACML 3.0, a well-known international standard for access control that defines that if any applicable policy returns deny, then that decision takes precedence over permit. This also runs counter to text anonymizers like Microsoft's Presidio, where redacted/detected spans are preserved. Indeed, it is very simply unfathomable that a privacy practitioner would adopt such a wasteful design that is leaky and allows redactions by a guard to be trivially invalidated by another guard.
  • [W3] Missing citations ([1,2]) for privacy guards involving input filters, output supervisors, etc., and mischaracterized related work ([3]); unlike the authors' claims in lines 049-050 and 099-100, [3] does not simply apply an edit or have single-site local individual guards, but offers a full suite of input and output guards and data minimizers. Additionally, while discussing privacy in multi-agent systems (line 107), they miss key prior work [4] on compositional privacy leakage in multi-agent systems that can circumvent redaction guards by composing responses to contextually allowed queries.
  • [W4] Very poor presentation, un(der)explained jargon, a lot of single-use notation that can be removed to simplify the presentation. Some non-exhaustive examples are:
    1. In lines 234-238, the authors do not clarify where they get the numbers in "Answer to puzzle (a)" from (as in is it a toy example or an empirical result), and the prose is confusing. The notion of a primitive guard is also underdefined. The authors, here and elsewhere, make an undefined reference to the appendix that makes it unclear why a reader needs to refer to that.
    2. Much of the notation like FJψ,ϕFJ_{\psi,\phi} is defined in an arcane, cascading manner that is hard to follow for a reader. DW(k,m;g)DW(k,m;g), FϕJ,ηd,TˉdF^J_\phi, \eta_d,\bar T_d are defined but not used in discussion/analysis. Additionally, the authors provide theorems without a plain text discussion of what it means and how it is of interest to a privacy researcher/reader in general. The presentation in this work is genuinely hard to follow and not for the lack of careful reading (which I have done a couple of times very thoroughly) and is not of a standard that a top conference paper should aspire to achieve. It would be beneficial to clear up redundant single-use notation, proofread the paper for readability, and explain the premise and utility of a theorem carefully.
    3. Terms like primitive calls, writer-reviewer loop, typed two-terminal guard circuits (what are the types?), primitive guards, pareto synthesis, etc. are not defined or underdefined. If defined in the appendix while being used in the main text, the authors do not mention where it is defined, and it is not good practice to use terms prior to their definition being provided.
  • [W5] The redaction guards used in their evaluation appear to be ad hoc and not compared against existing baseline guards. At the very least, their empirical results should clearly adopt/describe how they adopt the state-of-the-art guards for this setting, else this study may rely on flawed implementations, which undermines trust in their empirical claims.
  • [W6] While the authors introduce cost as a key variable to optimize, they never define how to measure it in practice/what quantity they seek to measure precisely (viz. units or a measurement procedure) nor are empirical results provided for cost measurement.

Additional Questions

  • [Q1] The paper is replete with claimed theorems and assumptions. Can the authors clearly map where/if they are empirically substantiated? Indeed, every empirical result that pertains to a theorem must be mapped to the corresponding theorem or assumption made. Additionally, each such result must employ a state-of-the-art sound implementation.

References

  • [1] Ghalebikesabi, S., Bagdasaryan, E., Yi, R., Yona, I., Shumailov, I., Pappu, A., Shi, C., Weidinger, L., Stanforth, R., Berrada, L., Kohli, P., Huang, P., & Balle, B. (2024). Operationalizing Contextual Integrity in Privacy-Conscious Assistants. ArXiv, abs/2408.02373.
  • [2] Das, S., & Fioretto, F. (2026). NeuroFilter: Activation-Based Guardrails for Privacy-Conscious LLM Agents. ArXiv, abs/2601.14660
  • [3] Abdelnabi, S., Gomaa, A., Bagdasarian, E., Kristensson, P.O., & Shokri, R. (2025). Firewalls to Secure Dynamic LLM Agentic Networks. ArXiv, abs/2502.01822.
  • [4] Patil, V., Stengel-Eskin, E., & Bansal, M. (2025). The Sum Leaks More Than Its Parts: Compositional Privacy Risks and Mitigations in Multi-Agent Collaboration. ArXiv, abs/2509.14284.
11 comments
AuthorsOpenReview ↗

Response 1/6

Thank you for your careful and detailed review of our paper. We are encouraged that you found our framing of guard composition as a circuit design problem intuitive and interesting, including the result that guarding the minimum cut roughly matches guarding all edges. Below we have provided responses to each of your concerns, which we hope will provide additional clarification.

The points under Critical Weaknesses are repeated under Questions For Authors with different numbering. Below, we follow the numbering under Questions For Authors (W1–W6 and Q1), which covers every point. Please let us know if there is any confusion.


W1. Is the framework only for redaction-based guards?

The framework is not restricted to redaction-based guards. Section 4.1 starts with a library of available guards, which can include deterministic rules, LM guards, human review, or hybrids. We call these primitive guards. Each is assigned four quantities: privacy leakage, task utility, latency, and cost. This definition does not depend on how the guard is implemented. Our experiments focus on LM-based guards.

Definition 3 introduces typed two-terminal guard circuits recursively. They are two-terminal because each complete circuit has an input and an output, and they are constructed from serial composition and parallel branches closed by a release switch. The types ensure that the connected inputs and outputs are compatible.

The guards in our experiments also remove information in different ways. The censoring guard deletes sensitive details, and the editing guard replaces them with typed placeholders such as [NAME1]. In the revised manuscript, we added two guards from the literature that do not redact (App. F.3). PrivacyChecker [1] analyzes each information flow under contextual integrity and replaces disallowed details with a high-level summary, i.e., it removes information by abstraction. Adversarial Anonymization (AA) [2] infers personal attributes from the message and passes these inferences to an anonymizer, which minimally edits the message to generalize or remove the cues behind them (e.g., "my husband and I" → "my partner and I"), i.e., it removes information by generalization. In addition, the release switch in App. F.9 is a classifier that decides which messages to release rather than editing them (see our response to W2). The composition trends are consistent across all of these guard types (Table 6), and the formal results do not depend on how a guard removes information.

AuthorsOpenReview ↗

Response 2/6

W2. Major flaw with parallel-any composition

Thank you for raising this point of confusion.

First, we want to emphasize that minimizing leakage is not the only goal. One of the core problems our work investigates is the balance between 1) privacy leakage, 2) task utility, 3) Latency, and 4) Cost. Specifically, the tradeoff between leakage and task utility has been well-established in other literature (e.g. [9, 10]). Previous works have similarly proposed allowing more potential leakage to improve utility. For example, Zhou et al. [11] start by redacting the sensitive spans detected in a prompt, and then relax these redactions step by step (first to an abstraction, then to the original text) until the utility of the response is preserved. The reasoning behind this approach is that existing guards are not perfect -- they can misclassify benign messages as sensitive and incorrectly block information that's necessary for downstream tasks.

Therefore, the parallel-any composition is intentional. In our experiments we explicitly show how the parallel-any composition navigates the tradeoff between task utility and leakage.

Second, the reviewer's described rule is already a part of the framework. In our formulation, parallel guards are optionally followed by switches that decide which of the outputs to release. Releasing all outputs is one option -- this is the option we use for the experiments to make the tradeoff between utility and leakage more clear. We discuss other release options in App. A.4. In particular, the rule the reviewer describes, where a message is released only if every parallel guard allows it, is the approval-quorum rule in App. A.4 with the quorum set to all guards. Our main composition theorem (Theorem 4) applies to any such release rule and is not limited to release-all.

Based on your feedback, we have added an experiment with a more conservative release switch (App. F.9 in the revised manuscript). We implement a zero-shot policy switch (ZS-Policy), adapted from the zero-shot baseline of LlamaGuard [6], which classifies each branch of a parallel censoring circuit as safe or unsafe and releases only the branches classified as safe. We are also currently running the finetuned LlamaGuard model and will update the results if they finish within the rebuttal period. We compare it with releasing all branches (Baseline (No Switch), i.e., the parallel-any rule used in Experiment 1) and with releasing one randomly selected branch (Baseline (Random Select)). The results at parallel width k=5k=5 are summarized below (Table 12 in the revised manuscript):

DatasetRelease ruleMessages blocked (%)Leakage ↓Task Success (Utility) ↑
ACIBBaseline (No Switch)0.00.1720.544
ACIBBaseline (Random Select)80.00.1210.450
ACIBZS-Policy20.60.1300.490
ASBBaseline (No Switch)0.00.1680.981
ASBBaseline (Random Select)80.00.0850.991
ASBZS-Policy4.60.1550.981
PAC-BenchBaseline (No Switch)0.00.8960.786
PAC-BenchBaseline (Random Select)80.00.8130.705
PAC-BenchZS-Policy37.90.7670.664

Messages blocked is the percentage of candidate branches not released.

ZS-Policy significantly reduces leakage relative to parallel-any on all three datasets, at the cost of lower task success on ACIB and PAC-Bench. This illustrates that the release switch is a design choice within the framework: a more conservative switch moves the parallel block toward privacy, while parallel-any moves it toward utility.

In summary, we investigate the parallel-any approach as one possible building block when composing guards, and investigate how incorporating it affects tradeoffs between 1) privacy leakage, 2) task utility, 3) Latency, and 4) Cost. We do not recommend this as a universal privacy policy for every situation; it is simply an optional part of the framework that generally improves task utility at the cost of leakage and can be combined with other release policies.

We will make this more clear in the manuscript to reduce confusion.

AuthorsOpenReview ↗

Response 3/6

W3. Related Works

Thank you for pointing out these references. In the revised manuscript we have added Ghalebikesabi et al. and NeuroFilter to our discussion of privacy guards, and added Patil et al. to our discussion of privacy in multi-agent systems (Section 2).

We have corrected our description of Abdelnabi et al., thank you for pointing this out.

These works propose individual guards, or a fixed arrangement of guards around one agent. Our work is complementary: we take such guards as given and study how combining them, and where they are placed in a larger workflow, changes leakage, utility, latency, and cost.

Finally, Patil et al. show that leakage can arise when individually harmless responses are combined. Our analysis tracks whether a given piece of sensitive information survives each guard, so it does not cover new inferences created by combining messages. We already state this limitation in App. G and will reference Patil et al. there.

Reviewer_XpD8OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper describes the composition of safety guards in the form of "circuits". In the abstract, it says the problem the paper tries to solve is "what behaviors and design principles emerge when imperfect guards are composed, and where should the resulting compositions be placed in an agent workflow?" The main sections of the paper describes some abstract notions of different types of circuits with some of its properties presented as theorems. There are some experiment results showing the utility and leakage tradeoff for different types of circuits.

Critical strengths
  • The "circuit" description of safety guards could be useful and interesting if the paper properly describes how it could be used in reality (which it currently doesn't describe clearly).
  • The paper has some experiments showing the success and leakage metrics when applying different kinds of circuits (although I don't understand what these results are trying to validate or why they are important).
Critical weaknesses
  • The paper provides a fancy description of different compositions of functions. It does not have a clear objective such as what is being maximized/minimized and under what constraints. Without a clear objective, it is hard to quantify what kind of configurations ("circuits") is good and what is not good.
  • It is unclear to me in what way the theorems in the paper are useful, because I do not understand what is the exact problem this paper is trying to solve.
  • The experiments only compare between different kinds of circuits. There appears to be no baseline methods being compared against. The usefulness of this work remains unclear.
  • The paper doesn't seem to really answer the question "what behaviors and design principles emerge when imperfect guards are composed, and where should the resulting compositions be placed in an agent workflow?" posed in the abstract. After reading the paper, I don't know how these circuits can be used in real agentic workflows.
  • My main impression after reading this paper is that it appears to solve a highly scientific problem but nothing can be understood by most reasonably tenured expert in the AI/ML field, so even if the results are valid, the paper may not be suitable for ICLR.
Questions for authors

I do not understand what real problem this paper is trying to solve, or how it solves the question posed in the abstract. Please refer to the weaknesses for details and address them if possible.

6 comments
AuthorsOpenReview ↗

Response 1/4

Thank you for taking the time to review our paper. We are encouraged that you found the circuit description of privacy guards potentially useful and interesting. Below we have provided responses to your questions, which we hope will clarify the objective of our work and the practical meaning of our theoretical results. Please let us know if you have any follow-up questions.


Q1. What is the optimization objective, and what constraints define a good circuit?

In our work we focus on guards, which are models that are placed on the communication edges in multi-agent systems for the purpose of screening messages, tool calls, etc, for leakage of private data. In particular we focus on stochastic and LLM-based guards. These guard models have various performance characteristics which have been investigated in the literature (e.g. [1, 2, 8]); we focus on 4: 1. Privacy Leakage (A), 2. Task Utility (U), 3. Latency (T), and 4. Cost (K). In general, it is desirable to minimize leakage, latency, and cost, while maximizing task utility. However, this is a challenging problem, as both where guards are placed and how they are arranged affect the overall performance.

We therefore separate the problem into the two design questions that structure the paper:

  • Paper Q1 (composition): How should guards be combined to optimize the 4 characteristics? (Section 4)
  • Paper Q2 (placement): Where should guards be installed? (Section 5)

To answer the composition question, we compose guards into circuits by explicitly formulating the optimization problem in Eq. (23) (Appendix C.4):

min⁡CK(C)\min_C K(C)

subject to

U(C) \geq \underline U, \quad T(C) \leq \bar T.$$ Here, we minimize circuit cost subject to privacy, utility, and latency constraints. The bounds depend on the task and the desired level of privacy and utility. We placed this formulation in the appendix due to space constraints, but we can bring it into the main body for clarity. For the placement question, Theorem 18 shows that the minimum-cost placement of perfect guards is given by a minimum cut, while Theorem 7 provides an upper bound on privacy leakage for imperfect guards using ordered barriers. To be clear, we are not claiming to build a better individual guard than existing methods; rather we take available guards and study how their arrangement and positioning affect these 4 factors. For example, under the stated assumptions, serial composition improves privacy by reducing harmful-information survival, while exposed parallel-any composition improves useful-information survival at the cost of additional leakage. Theorem 4, Corollary 5, and Experiment 1 explain this tradeoff. More generally, Theorem 6 and Corollary 17 justify a dynamic-programming approach to finding suitable circuits under a fixed budget.
AuthorsOpenReview ↗

Response 2/4

Q2. Why are the theorems useful?

Thank you for bringing up this point of confusion. The theorems help us understand to what extent our GuardCircuits formulation can be generalized to other compositions and multi-agent systems without having to empirically test against all possible systems. Below we give a more intuitive description of the key theoretical results, which we will update in the manuscript.

Paper Q1 (composition): How should guards be arranged? Suppose we are given a set of state-of-the-art guards. The first issue is that these guards are not perfect: for example they might accidentally leak private information or block useful information. We propose a circuit approach for composing these guards so that the overall design can achieve the desired privacy–utility tradeoff.

Theorem 4. For serial composition and parallel composition with a switch, we compute privacy leakage, useful-information survival (or utility), latency, and cost from the component signatures. These formulas use probability theory and require the stated assumptions, however they give a useful way to understand and predict the effect of an arrangement. The point is that we can compare architectures without treating every circuit as independent black boxes.

Building on these composition laws (Theorem 4), we have the following useful guidance for practical use:

Corollary 5. This gives a simple rule for balancing depth and width under a fixed number of guard calls in the high-reliability regime. More serial depth buys privacy; more exposed parallel width buys task reliability. The more general result with switches is Theorem 13 in the appendix. Experiment 1 tests the corresponding qualitative trends on real benchmark inputs and LM guards.

Theorem 6. This enables a dynamic-programming approach to circuit design. Suppose we arrange guards in a complicated circuit and later replace one small part with a better guard. Do we need to recalculate the entire system, or should the overall system improve? Theorem 6 shows that, under the stated assumptions, replacing a subcircuit with another that is no worse in privacy, utility, latency, cost, and blocking contract cannot hurt the full circuit. This is important for complicated systems and justifies pruning inferior candidates in the dynamic-programming approach of Corollary 17.

A more global question is where to place guards and what guarantees we can provide regarding privacy leakage in the whole system. This is the focus of Theorems 18 and 7.

Paper Q2 (placement): Where should the circuits be placed? Suppose we are given an agentic system with 10,000 agents. Where should guards be installed? Guarding every edge is expensive, and guarding the wrong edges may leave a path unprotected. We study this as a graph problem.

Theorem 18. For ideal blockers, the minimum-cost placement is a directed minimum cut between protected origins and unauthorized recipients. Minimum cut itself is a classical result; our point is to identify how it applies to this guard-placement problem. The placement experiments test whether this principle is also useful with imperfect LM guards. In fact, Experiment 2 shows that placing guards on only two minimum-cut edges achieves nearly the same leakage reduction as guarding all 17 edges. This demonstrates the practical usefulness of the theoretical placement principle.

Theorem 7. For imperfect (stochastic) guards, things are more complicated. We do not claim to find the globally optimal placement. Instead, we extend the minimum-cut idea to multiple ordered barriers in the execution graph and show that, under the conditional survival assumptions, leakage is controlled by an explicit probability upper bound. Such a guarantee is useful for understanding and designing systems even when exact leakage is difficult to calculate. The recurrent-cut experiment illustrates the role of repeated barriers when it comes to imperfect guards.

The theorem-to-experiment table under Reviewer P1jw, Q1 summarizes the usefulness of each theorem and its connection to the experiments.

AuthorsOpenReview ↗

Response 3/4

Q3. Where are the baselines?

The main point of GuardCircuits is not to design a new state-of-the-art guard, but instead how to compose and position existing guards. We can take existing guard or switch mechanisms and compose them (even with multiple copies of itself) to improve tradeoffs between leakage, utility, latency, and cost.

For Experiment 1, which evaluates guard composition, the baseline is simply each guard method by itself, i.e. k=1k=1. We then compare it with serial and parallel compositions of the same guard. As shown in Experiment 1, serial composition reduces privacy leakage compared with a single guard, while parallel composition can improve utility, showing the tradeoff between the two. We compare 2 different types of guards, censoring and editing, parametrized by 6 different LLM models. After the feedback from reviewer P1jw we have also added 2 more guard methods: PrivacyChecker [1] and Adversarial Anonymization (AA) [2]. This gives us 14 different guard models for comparison.

For Experiment 2, which evaluates circuit placement, we compare against baselines of having no guards, guarding all edges, and guarding random placements. We observe that using our minimum cut approach allows guards to achieve the same leakage and safe-task success rate with significantly fewer guard placements, achieving almost the equivalent of guarding all edges. The fewer placements results in significantly lower latency and cost for the system.


Q4. How would someone deploy GuardCircuits in an agent workflow?

Deployment depends on the task and the desired privacy, utility, latency, and cost constraints; we leave this general so that it can be adapted to a wide variety of applications.

For local circuit design, suppose one wants to protect a particular edge under certain privacy, utility, latency, and cost constraints. One can use our composition laws (Theorem 4) to understand how different arrangements affect these quantities, or use Algorithm 1 to find suitable circuits under a fixed guard-call budget. For instance, if privacy is more important, serial composition can help, while parallel composition can improve utility at the cost of additional leakage.

For a more general agentic workflow, one can first represent the system as a directed graph and identify the protected origins and unauthorized recipients. The next step is to decide where to place guards. For ideal guards, Theorem 18 gives the minimum-cut solution. For imperfect guards, Theorem 7 provides a privacy leakage upper bound when the placement creates ordered barriers satisfying the stated conditions. This can guide the choice of placement and the required privacy protection at each guarded edge.

Once the placement and desired privacy levels are determined, one can use the local circuit design approach above to find suitable circuits that satisfy the privacy constraints while preserving as much utility as possible under the cost and latency budgets.

Thus, the framework provides a principled way to move from global placement to local circuit design, rather than placing arbitrary guards at arbitrary edges.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.