acceptodds
Under review as a conference paper at ICLR 2027

Safe Online Personalization of Agent Permissions via Auditable Policy Distillation

Abstract

Personal AI assistants act with broad authority (reading files, sending email, executing code), while the permission policies that constrain them are written down in advance and fit no individual user. We formulate safe selective permission learning: an agent's governance layer must decide ALLOW/DENY/ASK per action online, learning each user's latent preferences from feedback that is costly to elicit and missing-not-at-random, while false allows on risky actions are budgeted per risk tier and every learned behavior remains inside immutable, user-declared policy ceilings. Our method couples a cost-sensitive Bayesian logistic learner (with ASK as an explicit observation action, inverse-propensity-corrected updates, and confidence-gated autonomy on high-risk tiers) with auditable policy distillation: learned decisions are periodically compiled into symbolic rules in a small policy grammar, admitted only after calibration-set validation, and enforced thereafter as inspectable, revocable policy. We instantiate the method as GUARD-OPENCLAW, a governance layer for the open-source OPENCLAW platform in which a deterministic trusted core validates typed, single-use, provenance-carrying capabilities at the tool boundary and writes hash-chained receipts ahead of execution. In a behavioral evaluation on the user-client side with template-grouped splits and six user personas, the learner reaches selective accuracy at autonomous coverage with a false-allow rate among autonomous allows, versus for linear Thompson sampling and for the best fixed policy (mined attribute rules), at a median end-to-end governance overhead of . The extension, benchmark generator, and a CPU-only harness that reproduces the learning dynamics of our method are released.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
agent security, permission control, policy distillation, auditable AI

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_Eb7DOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper tackles the problem of agents learning appropriate permissions for a given user in an online fashion. Modern agentic systems often make tool calls that are potentially unsafe (e.g. vulnerable to prompt injections or exfiltrating user secrets). As a layer of safe-guards, these systems can ask the user for permissions to call certain tools, at a usability cost. The present paper aims to learn a context-dependent policy specific to the given user which minimizes the usability cost as well as the cost of incorrectly allowing or denying a tool call in an online fashion.

Critical strengths

The formuation of the problem as an online learning problem is nice.

Critical weaknesses
  1. One of my major concerns is with the writing of the work. The paper is extremely hard to read as a human. The phrasing is awkward, the sentence structure is hard to follow, some symbols seem undefined (e.g. ρ\rho in Eq (1), IPW in the results) . The method section is also very poorly written, e.g. what is the channel oo, what is WW in computing the propensity, how are μ,H\mu, \mathbf{H} determined. The system design does not seem to be well motivated, or at least the motivating principles are not well articulated. I also do not understand when one asks, the expected costs seem to imply that p_{ask}=1 ?
  2. The method is also extremely complicated with a lot of moving parts, and it would be good for the paper to work through an example to clarify how the method works. I also do not understand why each component of the method is useful, e.g. why is policy distillation done every few steps? What happens if this were not incorporated in the system? These are also not empirically demonstrated through ablations.
  3. The objective in Eq (1) is not well motivated. Notably, the cost should be assymetric (i.e. a false allow is often catastrophically bad from a security perspective, despite Φ\Phi), the cost for asking needs to change with the number of asks
  4. Critical prior work is missing -- https://arxiv.org/pdf/2511.17959 proposes using a similar learning based approach to asking for user permissions.
  5. There is no human evaluation. This seems to be crucial.
Questions for authors

Apart from the weaknesses above,

  1. Could the authors provide a motivation (intuitive or empirical) for all the moving parts of the system?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Eb7D

Thank you for the specific criticisms on writing, symbols, motivation, and the missing citation. We address each below.

  1. Writing and undefined symbols (your first weakness)

Reviewer's point: The paper is extremely hard to read; some symbols seem undefined (Eq. (1), IPW in the results); the method section is poorly written (what is the channel, how is the propensity computed, how are the thresholds determined). Our response: A full readability rewrite was done. Every quantity of Eq. (1) is now defined in place (cost-matrix semantics and magnitudes, delta, the conditional meaning of FAR); the four feedback channels and their propensities are tabulated (Appendix A); IPW is expanded at first use; the thresholds theta_f/theta_a/n_min/kappa_h/k_o are defined where they are introduced; and the IPW weights are given exactly (1.18/1.03/2.86).

  1. The ASK arm and p_ask (your second weakness)

Reviewer's point: The expected costs seem to imply p_ask = 1 when the system asks. Our response: Clarified: E[cost | ask] = c_ask + (1 - pi_a) c_ig, where pi_a = 0.85 is the probability the user answers. The second term is round-independent, so it never changes which action minimizes the cost, but it belongs in the cumulative interaction-cost accounting; an answered ask incurs no decision-error cost, and utility counts a task only if it completes autonomously or after an answered ask. The original text wrote only c_ask, which created the confusion — thank you for pointing it out.

  1. Worked example and component motivation (your third weakness)

Reviewer's point: A worked example would clarify the many moving parts; the usefulness of each component (e.g., distillation every few steps) is not demonstrated.

Our response: A worked example was added (Appendix B): one concrete request (a 21:00 critical send) is traced through p~ = 0.55, the three arm costs (5.4/0.55/0.37), the gate test (1 - p_lo = 0.10 > 0.01, hence ASK), the update weight 1.18 after the user's answer, and the next distillation cycle with the fit/calibration fidelity and false-allow arithmetic, admission, and user confirmation. A new paragraph maps each component to the failure mode it removes, with ablation evidence. Please also note that Table 2 already contained four distillation ablations (w/o distillation, cell-constant, w/o calibration validation, w/o demotion); we additionally swept the period T_d: {100, 250, 500} are indistinguishable; 1,000 degrades; never distilling gives 61.2% coverage and 3.8% FAR.

  1. Motivation of Eq. (1) and the ask cost (your fourth weakness)

Reviewer's point: The objective is not well motivated; the cost should be asymmetric, and the ask cost should change with the number of asks.

Our response: The cost matrix semantics and magnitudes are now stated (1/2/6/12; fd = 1; ask = 0.35; ig = 0.15), with the asymmetry motivated (a critical false allow is an irreversible external effect: a wire, a public post, credential disclosure); the budgets delta follow from cost-normalized risk (0.05 x 6 = 0.30 vs. 0.01 x 12 = 0.12). New ask-fatigue experiment: the per-ask cost is multiplied by (1 + gamma x asks so far) for gamma in {0, 1e-4, 5e-4, 1e-3}. Measured: asks per 100 requests fall from 9.3 to 5.4, coverage 90.7 -> 94.6, SelAcc 97.0 -> 94.6, utility 97.0 -> 92.7, FAR 2.5 -> 2.8 — the learner front-loads its interruption budget and then guesses more; the flat cost (gamma = 0) is the best operating point.

  1. Missing prior work (your fifth weakness)

Reviewer's point: Critical prior work is missing: arXiv:2511.17959 proposes a similar learning-based approach to asking for permissions. Our response: Thank you for the pointer. We verified and now cite Wu et al., Towards Automating Data Access Permissions in AI Agents (IEEE S&P 2026; arXiv:2511.17959). Their user study (205 participants; 85.1% accuracy, 94.4% at high confidence) actually strengthens our motivation (preferences vary with context but are consistent within a context). Our four deltas: online MNAR learning (vs. a static history); asymmetric, per-tier false-allow budgets; distillation into enforced symbolic rules; and the interruption budget as a first-class objective (enforcement is explicitly out of scope in their work).

  1. Human evaluation (your sixth weakness)

Reviewer's point: There is no human evaluation, which seems crucial.

Our response: we cannot complete a user study within the rebuttal window (institutional review and recruitment take weeks), and we will not fabricate one. We have pre-registered a two-stage protocol in the paper (Appendix J): Stage 1, a usability pilot (N ~ 20; rule readability, time to locate a rule, error discovery on three seeded mis-distilled rules, SUS); Stage 2, personalization vs. interruption (N ~ 100, three randomized conditions, mixed-effects analysis). We commit to completing Stage 1 before camera-ready. All current claims are scoped to synthetic personas.

Thank you for your response. Given that the entire paper is rewritten it needs another round of review to assess its merits. I hence maintain my current recommendation.

Reviewer_5QRuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a system that learns a user's permission preferences over time. It decides whether to allow an action, deny it, or ask the user. It also learns permission rules from feedback. These rules can be checked and removed, while the user's fixed restrictions stay in place.

Critical strengths

The paper addresses a practical problem, and the system design is fairly complete. A strength of the design is that it combines fixed user rules with rules learned over time, with the user's fixed rules having higher priority. The experiments include several baselines and ablations, and the reported results show fewer incorrectly allowed actions while keeping most decisions automatic.

Critical weaknesses

The paper should make its new contribution clearer, especially in the policy distillation step. The rules seem to be learned directly from user feedback. The authors should clarify how the learned preference model is used to generate these rules. They should also explain how this differs from existing rule learning and policy distillation methods, and what advantages their approach offers.

The experiments use synthetic users whose preferences and feedback closely match the method's assumptions. For example, the learner is given the true probabilities of observing feedback, which may be hard to estimate in practice. It is unclear whether the same gains would hold when these probabilities are estimated poorly, or when users have more complex preferences or less consistent feedback.

Questions for authors

See the Critical Weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 5QRu

Thank you for the positive assessment of the design and evaluation, and for the two criticisms — both are now addressed.

  1. The distillation contribution and how rules are generated (your first weakness)

Reviewer's point: Clarify how the learned preference model generates the rules, and how this differs from existing rule learning and policy distillation.

Our response: Section 3.3 was rewritten to make the pipeline explicit. The statistical learner contributes to rule generation in three ways: (i) its ASK decisions concentrate labels on the cells that are actually ambiguous, so evidence accumulates where rules are needed; (ii) its silent decisions determine which cells have override and confirmation evidence at all; (iii) cells whose boundary the grammar cannot express remain statistical. The rules themselves are fitted on the logged labels (not extracted from model weights) and are admitted only after calibration-set validation against three thresholds (fidelity theta_f, false-allow contribution theta_a, support n_min), with a per-cycle cap, provenance metadata, individual revocation, and demotion after k_o consecutive overrides. Difference from existing work: ABAC policy mining and AgentGuardian-style learning fit rules offline, once, with no feedback loop and no validation gates; classic policy distillation (new citations: Rusu et al. 2015; Hinton et al. 2015) compresses one network into another — we distill into a human-readable policy language whose output a deterministic core enforces, so the artifact is auditable rather than merely a smaller model. New ablations: a rules-only layer (no statistical decisions; uncovered cells always ask) reaches comparable converged metrics on the 36-cell harness (Cov 90.7, SelAcc 97.6, FAR 2.4) but its coverage at round 250 is only 6.5% (everything asks until rules accumulate), and on the open-ended cell space of a real platform most cells would never reach the support threshold under such a policy — that is the role the statistical layer plays. Distillation-period sweep: T_d in {100, 250, 500} are statistically indistinguishable; 1,000 costs coverage (84.5); never distilling costs 61.2% coverage and 3.8% FAR (the time-conditional cells can only be automated by the distilled time-window rules).

  1. Synthetic users match the assumptions (your second weakness)

Reviewer's point: The experiments use synthetic users whose preferences and feedback closely match the method's assumptions. Our response: The same new experiments as for Reviewer 2, item 3: audit-anchored propensity estimation matches the oracle, RFF/MLP nonlinear ground truth does not degrade the learner, and an ask-fatigue experiment models count-rising interruption costs (gamma in {0, 1e-4, 5e-4, 1e-3}). We still scope all claims to synthetic personas and commit to the pre-registered user study (Appendix J).

Reviewer_Wx64OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies online personalization of agent permissions through ALLOW/DENY/ASK decisions under selective feedback, asymmetric error costs, and immutable policy ceilings. The proposed method combines Bayesian logistic preference learning, inverse-propensity weighting, randomized audits, and risk-dependent confidence gates. It periodically converts observed behavior into symbolic rules subject to calibration checks and revocation. The Guard-OpenClaw implementation adds typed single-use capabilities, provenance checks, and write-ahead receipts. Evaluation uses six synthetic personas, template-grouped queries, component ablations, drift scenarios, and adversarial tests. The paper reports 96.9% selective accuracy at 91.4% autonomous coverage, with a 1.8% false-allow rate and approximately 0.44 seconds of governance overhead.

Critical strengths
  1. The paper addresses a relevant and practically important problem. Personalized permissions must balance user interruptions against asymmetric authorization errors, and the paper makes these competing objectives clear.

  2. The rule grammar, provenance metadata, validation checks, and demotion mechanism produce explicit artifacts that users can review and modify. The learned policies are inspectable and reversible.

  3. Comprehensive experiments are done to evaluate the model from different perspectives. Template-grouped splits and a shared enforcement stack help isolate learner effects, while the evaluation examines cold start, drift, poisoning, and attacks alongside decision accuracy.

Critical weaknesses
  1. The conformal guarantee does not match the stated safety objective. Lemma 1 bounds the joint probability of an erroneous allow, whereas Eq. (1) constrains error conditional on allowing within each tier. Appendix J additionally calibrates pooled audits at 5% for both budgeted tiers, which does not establish the critical tier’s 1% conditional-risk guarantee.

  2. Table 2 reports that removing IPW increases FAR from 1.8% to 2.3%, whereas Appendix L reports that the released harness produces the opposite direction, from 1.88% to 1.65%. The results seems abnormal and need to be further investigated.

  3. The simulation strongly favors the learner’s modeling assumptions. Appendix L uses linear persona scorers and supplies exact feedback-channel propensities, leaving robustness to preference-model misspecification and estimated propensities unestablished.

Questions for authors
  1. What guarantee is intended for the conformal variant: joint error control or the per-tier conditional FAR constraint in Eq. (1)? Please explain how the 1% critical-tier target is justified.

  2. Does the method retain its advantage when observation propensities must be estimated and preference boundaries are not generated by linear scorers?

  3. Given the reported feedback burden, under what usage scenarios does the method offer a practical advantage over simpler permission policies? How does the reduction in interruptions balance the user effort required for personalization.

  4. How well does the approach generalize to different agents and tool environments? Specifically, which components can be reused directly and which require additional training or manual configuration.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Wx64 (1/2)

Thank you for the careful review. The IPW inconsistency you caught is a real error; we investigated and corrected it (item 2).

  1. Conformal guarantee vs. Eq. (1) (your first weakness and Q1)

Reviewer's point: The conformal guarantee does not match the stated safety objective; the pooled 5% calibration does not establish the critical tier's 1% conditional guarantee. What guarantee is intended, and how is 1% justified?

Our response: You are exactly right. Section 3.2 now states that the conformal variant's intended guarantee is the per-tier conditional constraint of Eq. (1), FAR_rho^auto <= delta_rho; Lemma 1's joint bound is the intermediate step, and dividing by the admission probability yields the conditional form. The pooled 5% calibration bounds only the pooled conditional FAR — it does not establish the critical tier's 1%; our earlier wording was unclear and has been corrected. To address the question directly, we implemented and measured per-tier calibration (each tier calibrated on its own audit labels at its own delta with the (n+1) correction; new table "Per-tier conformal calibration"). Measured on the server (3 seeds x 10,000 rounds): the pooled gate opens at rounds 3,684–4,019 (19 labels is only the theoretical all-allow minimum; the measured opening is later); the per-tier high gate opens at 2,579–4,793; the critical gate does not open within 10,000 rounds — the learner's critical-tier audit sample (85 labels, 80% denies) has no score prefix below a 1% deny rate. The strict 1% conditional guarantee is thus achieved by failing closed: no autonomous critical allow is admitted. This is the honest statement: unless the learner's confidence tail on a tier is genuinely sub-1%, a 1% conditional guarantee and autonomous critical allows are incompatible. The deployed design already layers the fallbacks: ceiling-designated critical operations always ask; distilled critical ALLOW rules require explicit user confirmation; and the LCB gate is the model-based approximation for the residual critical cells. The 1% budget is now justified in Section 2.2 by cost-normalized risk equalization: 0.05 x 6 = 0.30 (high) vs. 0.01 x 12 = 0.12 (critical); the critical tier is held stricter per unit cost because its errors are irreversible and externally visible.

  1. The IPW ablation contradiction (your second weakness)

Reviewer's point: Table 2 reports that removing IPW raises FAR from 1.8% to 2.3%, while Appendix L says the released harness produces the opposite direction (1.88% -> 1.65%).

Our response: Thank you for catching this — it is a real error on our side. The submitted Table 2 value (2.3%) was an estimate that the released harness does not reproduce. We re-ran the ablation with the released code, unchanged, on the server node (10 seeds x 2,000 rounds). Measured (IPW vs. no-IPW): Cov 91.3 +/- 5.7 vs. 90.6 +/- 6.9; SelAcc 97.3 +/- 1.2 vs. 97.8 +/- 0.7; FAR^a 2.1 +/- 0.7 vs. 2.0 +/- 0.8; unauthorized executions 18.1 vs. 13.2 per 2,000 rounds. Mechanism: on silent decisions the override channel fires on mistakes (0.97) far more often than the confirmation channel fires on successes (0.35), so the uncorrected posterior over-weights the minority label on each silent cell — a conservative shrinkage away from the learner's own decisions. IPW removes that shrinkage and moves the posterior to the unbiased boundary; it buys coverage and automation, and its effect on FAR is regime-dependent (both signs appear in the new robustness table). The Table 2 row and Appendix L now report the measured values, and the submitted estimate is retracted.

  1. Estimated propensities and nonlinear boundaries (your third weakness and Q2)

Reviewer's point: The simulation favors the modeling assumptions (linear scorers, true propensities). Does the method retain its advantage when propensities must be estimated and preferences are nonlinear?

Our response: Two new experiment groups (new table in Appendix L). (a) Propensity estimation: the learner receives no propensity values and estimates each channel's propensity online from the randomized audit stream (Laplace-smoothed empirical rates). Measured: SelAcc 97.4 vs. 97.0 for the oracle, FAR 2.3 vs. 2.5 — estimation matches the oracle within noise; severe misspecification (confirm x2, x0.5; override 0.50) costs at most 1.1 SelAcc points and never breaks a tier budget. (b) Nonlinear preferences: the persona's ground truth is a random-Fourier-feature or two-layer-MLP scorer while the learner remains a linear logistic model (i.e., misspecified). Measured: SelAcc 97.5/97.3 vs. 97.0, FAR 2.3/2.4 vs. 2.5. The advantage does not depend on the linear-scorer/known-propensity assumptions.

AuthorsOpenReview ↗

Response to Reviewer Wx64 (2/2)

  1. Practical advantage (Q3)

Reviewer's point: Given the feedback burden, when does the method offer a practical advantage over simpler policies?

Our response: Quantified: over 2,000 rounds the user answers about 430 prompts (about 41 user-minutes at 6 s each), versus 11 person-hours to author a static ACL that still mis-allows 15.9–24% of risky requests. The advantage materializes when requests concentrate on a recurring but non-enumerable cell space, when preferences are context-dependent (time, scope, recipient), and when false allows are expensive enough to justify the interaction budget. For small fixed tool sets, a hand-written ACL remains the right tool — and it is exactly SSPL's ceiling layer; the learner only fills the cells it leaves open.

  1. Generalization and reusable components (Q4)

Reviewer's point: How well does the approach generalize to other agents and tool environments?

Our response: A component-reuse table was added (Appendix G): reuse directly — the learner, gates, distillation, policy grammar, and receipt ledger (all platform-agnostic); adapt — five platform interfaces (descriptor source, policy decision point, capability issuer, execution hook, receipt sink) plus one canonicalizer per tool class; train — only the intent model, and only if the new platform has none.

Reviewer_jCe7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Agentic systems work under fixed static settings, and the paper's goal is to learn a policy to adapt such policy to downstream user using such system. The main challenge identified to adapt this policy to the user is their feedback is selective and missing not at random. The paper proposes some cost sensitive learning mechanism to model actions such as allow/ deny / ask.

Critical strengths

The problem seems to be interesting. It would be useful to adapt adapt agentic systems to the user. However, my evaluation of the work is severely hindered by the writing quality (more on this below).

Critical weaknesses

The biggest weakness for me is clarity and mis-attribution. Both stems from the AI usage. The paper AI use statement claims AI was used for "grammar revision," however Claudish markers are present starting right from the abstract with undefined terms and non-standard terminologies, making it difficult for me any legible reading of the work. As a reviewer, I understand that AI usage is allowed, however the requirement is also that the AI usage should be honestly written in the AI usage statement. Unfortunately, the paper seem to be violating this.

Beyond this, the clarity issue prevents me from making sense of the paper. As such, my recommendation is rejection with the reasoning that the paper's are written for scientific community, and as such should be legible to the broader audience.

Questions for authors

None

1 comment
AuthorsOpenReview ↗

Response to Reviewer jCe7

Thank you for the candid assessment. We apologize for the readability problems, and we treated the rewrite as a top-priority, whole-paper effort rather than a patch:

Terminology and readability

Reviewer's point: Undefined terms and non-standard terminology appear starting from the abstract. Our response: Every abbreviation (SSPL, MNAR, IPW, FAR, LCB, Cov/SelAcc/Util, R80) is now expanded at first use; internal terminology (authorization cell, policy ceiling, undecided cell, distilled rule) is unified and collected in the notation table; informal expressions such as "load-bearing", "daggers", and "one-click" are removed. The abstract and Sections 2–3 were rewritten so that each term is explained in place; the four feedback channels and their propensities are tabulated (Appendix A), and a fully worked numeric example of one request through decision and distillation was added (Appendix B).

AI-use statement

Reviewer's point: The AI-use statement claims only grammatical revision, which does not match the visible AI drafting.

Our response: You are right. The original statement disclosed grammatical revision but not the AI-assisted drafting and restructuring of prose — our disclosure was incomplete. The statement now lists four uses, explicitly including drafting, restructuring, and polishing the manuscript prose (including terminology normalization and readability revision) under author direction; it states that all reported numbers were produced by executing the disclosed code and that every AI-drafted passage was line-edited and approved by the authors. Thank you for enforcing this policy point.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to all reviewers

We thank the reviewers for careful and constructive feedback.

We revised the manuscript accordingly; every change is marked in blue in the revised PDF. All supplementary experiments were executed with the released code on a shared compute node (Ubuntu 22.04.4 LTS; two AMD EPYC 7542, 128 threads; 692.7 GiB RAM; eight NVIDIA RTX 6000 Ada 48 GB. Each new measured table states its replication command.

Revision overview:(1) A full readability rewrite: every abbreviation is defined at first use, terminology is unified, non-standard phrasing is removed, and the AI-use statement is now accurate. (2) The IPW ablation is corrected (Reviewer 2): the submitted Table 2 value (2.3%) was an estimate not reproduced by the released harness; it has been replaced by measured values with a mechanistic explanation. (3) New robustness experiments: propensity estimation and misspecification, nonlinear personas (RFF/MLP), per-tier conformal calibration, ask fatigue, distillation-period sweep, and a rules-only ablation. (4) New content: a worked example with concrete numbers (Appendix B), a component-reuse table (Appendix G), a pre-registered user-study protocol (Appendix J), and three new citations (Wu et al., S&P 2026; Rusu et al. 2015; Hinton et al. 2015). (5) Guarantee clarification: the conformal variant's intended guarantee is now stated as the per-tier conditional FAR of Eq. (1), and the 1% critical budget is justified by a cost-normalized derivation.

We thank all reviewers again for their time and feedback. We believe the revision addresses every point raised; we welcome further comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.