acceptodds
Under review as a conference paper at ICLR 2027

Auditing Write-Path Privacy in Shared Agent Memory: Source Counterfactuals and Authorized Sharing

Abstract

We audit how private messages enter shared agent memory and implement a restricted release interface for useful collaboration. A scheduling task may authorize an absence to be shared without authorizing its reason. Matched source replacements and non-injected questions test this boundary. On 16 fixed personas, judge-detected borderline-fact exposure is 0.875 for profiles and 0.594 for digests; model-assisted adjudicated reference rates are 0.922 and 0.719. Independent human reannotation retains a pooled profile-minus-digest contrast of 0.094 over 48 personas. Software cases trace storage, retrieval, and publication: five of twenty audited open-source workspace assistants exhibit the unbound configuration as shipped or by default, and two assistant deployments merge private messages into shared state in their tested defaults. An openclaw pilot exposes a private target in 3/4 channel-requested digests against 0/4 source-placebo digests. In 128 scheduling trajectories across four configurations, the two shared-workspace configurations let protected names reach a reader holding only the saved digest, while both audience-separated configurations show no detected disclosure on complete-evidence units. Native isolation already supports useful sharing, and an owner-confirmed release interface is not established as superior. External checks (the full public MuPPET and AgentLeak releases and an official-code CI-Work reconstruction) find AgentLeak leakage at 0.845 unbound and 0.441 under a prompt policy, reduced to 0.023 by a trusted pre-consolidation scope. The contribution is a reproducible audit and executable task-specific configuration evidence, not a general defense; simulated transport, ideal owner approval, and automated labels limit deployment claims.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
privacy audit, agent memory, access control, information leakage, LLM agents

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_QwWtOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates privacy leakage during the consolidation of private conversations into shared agent memory. The central concern is that private information may be incorporated into team-visible profiles or digests even when query-time access controls are correctly enforced. The authors construct a synthetic workplace benchmark and use matched source counterfactuals to examine whether private facts propagate into shared artifacts. They further evaluate exposure under different memory templates, privacy instructions, and ordinary user queries. Finally, the paper compares several configurations for authorized information sharing, including native agent isolation and an owner-confirmed release interface. The results demonstrate substantial privacy exposure in configurations without structural source binding, while audience isolation and trusted source filtering substantially reduce the observed leakage.

Critical strengths
  • The paper examines an important privacy boundary in multi-user agent systems: private information can enter shared artifacts even when subsequent access controls operate correctly. The non-injected threat model is particularly relevant because leakage can occur through ordinary collaborative interactions.
  • The authors conduct matched source counterfactual experiments, extend the benchmark to 48 personas, examine multiple writer and judge models, and include independent human reannotation. The LOFO controls provide useful evidence that observed disclosures depend on the private-source content.
  • The Mem0, LettaBot, and OpenClaw case studies complement the synthetic benchmark and demonstrate that the audited configurations are not purely hypothetical. The authorized-sharing experiments also highlight the practical
Critical weaknesses
  • Limited novelty beyond existing contextual-integrity research. The central privacy risk substantially overlaps with prior work, particularly PiSAs and Collaborative Memory. While the paper adds source counterfactuals and software case studies, it remains unclear what fundamentally new failure mode or diagnostic capability it establishes beyond existing studies of cross-user privacy leakage and shared agent memory.
  • Confounded primary comparison and unclear mechanism. The Profile–Digest comparison simultaneously changes organizing goals, field semantics, and person-slot counts (4 vs. 2). The borderline facts also closely match Profile's scheduling-related fields. Although additional factorial controls are provided, they do not consistently isolate the contributing factors. Thus, the primary result demonstrates a difference between two specific templates rather than a generalizable insight into memory organization.
  • Unclear technical contribution of write-path auditing. LOFO establishes that leakage depends on private-source content but does not identify where privacy restrictions fail. The main benchmark collapses memory writing and consolidation into a single summarization step, while the real-system case studies examine different stages without a unified evaluation. Consequently, the proposed audit appears difficult to distinguish from conventional privacy-sensitive summarization and source-ablation testing.
  • Unproven benefit of Confirmed Release. Native Isolation already achieves no detected privacy disclosure on complete-evidence trajectories. Confirmed Release improves joint success from 26/32 to 30/32, but the difference is not statistically significant and relies on an idealized owner. The paper does not demonstrate what additional failure mode this mechanism prevents compared with native isolation.
  • Sensitivity of the primary findings to evaluation methodology. The Profile–Digest contrast varies considerably across evaluators. Independent human reannotation reduces the pooled difference to 0.094, and the extension cohort's confidence interval includes zero. Documented confusion between scheduling constraints and private reasons further raises concerns about annotation reliability and the robustness of the primary conclusion.
Questions for authors
  • What new failure mode or diagnostic capability does write-path auditing offer beyond existing contextual-integrity research (e.g., PiSAs)? Given that the main benchmark collapses writing and consolidation, how does LOFO provide stage-specific insights beyond conventional source-ablation tests?
  • How can the Profile–Digest difference be interpreted when multiple factors (field semantics, organizing goals, and slot counts) change simultaneously? Moreover, how do the authors reconcile the substantially different estimates from automated and human evaluations?
  • What additional security benefit does Confirmed Release provide over Native Isolation, given that both show no detected leakage on complete-evidence trajectories and the improvement in joint success is not statistically significant?
1 comment

Your main point is fair: the spillage phenomenon is not new, and Confirmed Release does not add a proven security benefit. We have narrowed the claims in the revised paper.

PiSAs, Collaborative Memory, and LOFO. PiSAs already studies non-adversarial spillage across shared memory and derived notes, including availability versus private reasons and paraphrase-tolerant evaluation. Collaborative Memory already provides private/shared tiers, immutable provenance, and read/write policies. We do not claim either the spillage phenomenon or provenance-based isolation as a new discovery.

You are also right that LOFO is a source-ablation control, not a stage-specific diagnostic. Trellis merges memory writing and consolidation into one transcript-to-artifact step, and the Mem0, LettaBot, and OpenClaw cases examine different system boundaries rather than one unified pipeline. LOFO therefore supports the private source's contribution to disclosure. Fixed-transcript controls also show that assistant replies can retain private facts. It does not show which internal stage failed.

Within that narrower scope, our increment is the combination of matched source counterfactuals, separate measurement of artifact inscription and recovery by an independent artifact-only reader, annotation error analysis, and executable runtime configuration comparisons.

Profile–Digest. The comparison bundles organizing goal, field titles, and person-slot count (4 vs. 2), and all borderline facts are scheduling-related. Our factorial and capacity results are not stable across judge families. The revised paper therefore describes a difference between these two concrete templates on these scheduling assets. It does not claim a general person-centered/event-centered design rule.

Annotation. On the 48-persona matrix, the pooled Profile-minus-Digest contrast is 0.229 under the automated judge, 0.260 under model-assisted adjudication, and 0.094 under returned independent-human labels. On the 32-persona extension, the returned-human estimate is 0.055 (95% CI [-0.008, 0.117]), which includes zero and does not independently establish the contrast. The targeted review gives 0.255, but 174 of the 181 reviewed items were returned-positive by construction, so it cannot estimate residual error across a representative positive-and-negative sample.

The substantive disagreement is whether an artifact that records only an availability constraint also discloses its omitted private reason. We report the protocols side by side rather than choosing the largest estimate as the preferred one.

Confirmed Release. We agree that it demonstrates no additional security benefit over Native Isolation. Native already achieves zero detected disclosure on all 29 complete-evidence runs. Confirmed Release's 30/32 versus 26/32 joint-success difference is not statistically significant and relies on an idealized deterministic owner.

Across the 128 runtime trajectories, 115 have complete evidence; the other 13 are 3 workflow failures, 8 generation timeouts, and 2 unjudged runs, none treated as privacy-safe. In the revised paper, Native Isolation is the primary effective baseline, and Confirmed Release is retained only as an executable configuration comparison.

Reviewer_5adsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary
  • The paper attempts to study the problem of privacy leakage in memory that could be shared across agents in a team.
  • The threat model in the paper assumes that agents create some artifacts based on their past interactions and those artifacts could contain privavate information that is not meant to be shared with other agents. However, the threat model in the paper assumes that these artifacts are shared with other agents which causes accidental leakage.
  • The paper studies how memory organization affects this leakage and whether sharing information only when authorized can help prevent it.
Critical strengths
  • The paper evaluates across its own synthetic benchmark as well as other benchmarks such as AgentLeak and across open-source memory systems such as openclaw and mem0.
Critical weaknesses
  • The paper is difficult to follow
  • Terms such as "structural source binding", "unbound configuration" and "authorized projection" from definitions with concrete examples in the paper.
  • The threat model assumes that all the artifacts that the agent creates are shared with other agents in the team which is questionable. So the paper's main memory configuration doesn't have a private artifact which could bias the agent to store the information in the shared memory. -In Sec 4 and 5, the research qquestions that the experiments are trying to answer are not clear. There are multiple benchmarks, interventions, annotation methods which are intertwined and the main takeways are not clear.
  • The native isolation baseline is already reducing the privacy leakage. So the `owner confirmed release interface' does not add any value.
Questions for authors
  • Under what realistic deployment scenarios would an agent put a personal message in a shared artifacted rather than putting only the part which it is authorized to share in the shared artifact and the rest in its private memory?
1 comment

Thank you for asking when this situation occurs in practice. Here is the simplest example.

An employee privately tells the assistant, "I have an oncology appointment Tuesday 2–4 pm, please block my calendar for the team sync." The assistant may tell the team when the employee is unavailable, but not why. The original DM stays private; the derived team digest does not have to contain every detail from it. Correct read access to that digest does not make every copied fact authorized for teammates.

Our openclaw pilot tests this private-DM-to-channel-digest path. The judge finds the private target in 3/4 full-source digests and 0/4 source-placebo digests. The authorization and runtime experiments test the narrower case where availability may be shared but the reason may not. The 5/20 open-source audit is a convenience sample showing that such unbound default configurations exist in shipped code, not an estimate of deployment prevalence.

A real limitation of the main harness. Your criticism is fair: Trellis gives the writer only the shared template and no separate private derived-artifact store. That isolates how much private content enters an unbound shared artifact, but it can also push the writer to record facts there. We did not run a hybrid writer with both private and shared derived stores, so the revised paper limits the main-harness conclusion to this unbound configuration.

What the terms mean in plain language.

  • Source binding: which inputs may contribute facts to a given output—for example, whether a private DM may feed a team profile.
  • Unbound configuration: private and shared inputs are merged into a team-visible artifact without a structural source restriction.
  • Authorized projection: the part that may be shared more widely, such as Tuesday 2–4 pm unavailability without the medical reason.

The revised Sections 4 and 5 are organized around four questions:

  1. Does private-source content enter shared artifacts? Matched source replacements (LOFO) test whether the disclosures trace back to the private DMs.
  2. Can a teammate recover it from the artifact alone? An independent reader sees only the saved artifact.
  3. Does the template change exposure? Profile exceeds Digest on our scheduling assets, but the size depends on the annotation protocol and is specific to this template pair.
  4. Which runtime configuration supports useful authorized sharing? Audience separation works; shared-workspace configurations leak protected names into saved digests.

Native Isolation versus Confirmed Release. We agree with your assessment. Native Isolation has 29 complete-evidence trajectories with no detected disclosure; its task success is 29/32 and its joint success is 26/32. Confirmed Release has 30 complete-evidence trajectories and is not statistically superior, so it shows no established additional security benefit.

Across the full 128-trajectory matrix, 115 trajectories have complete evidence. The other 13 are 3 workflow failures, 8 connection timeouts, and 2 unjudged runs; none is counted as a privacy negative. In the revised paper, Confirmed Release is presented only as a task-specific configuration comparison, not as a superior defense.

Reviewer_aZn8OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of whether access control on shared agent memory is sufficient to preserve the privacy restrictions of the conversations from which that memory was derived. To put this in a simple question, the work asks whether a teammate can recover private information simply by reading a shared artifact that they are legitimately authorized to access. To study this problem, the work distinguishes source access, destination audience, and authorized release, and introduces an audit that combines matched source replacements with separate measurements of artifact exposure and reader recovery. The initial Trellis experiment compares profile and digest templates over 16 fixed personas using two writer models. The results show judge-detected borderline-fact exposure rates of 0.875 for profiles and 0.594 for digests, while independent human reannotation retains a smaller pooled profile-minus-digest contrast of 0.094 over 48 personas. The paper further examines memory frameworks, assistant deployments, external benchmarks, and 128 scheduling trajectories, comparing four runtime configurations. The runtime results show that audience isolation can support useful sharing, while owner-confirmed release is not established as superior to native isolation. As the work states, the contributions are (1) a source-counterfactual audit of private information entering shared artifacts, (2) measurement checks that distinguish inscription, recovery, and annotation uncertainty, and (3) executable configuration evidence for task-specific authorized sharing.

Critical strengths

The work has three strengths at least: (1) a clearly privacy boundary with useful source-attribution controls, (2) substantial measurement checks and careful statistical treatment, and (3) an executable evaluation of useful authorized sharing. First, the paper clearly separates permission to read a source from permission to publish information derived from it. This distinction is particularly useful for partial authorization, where an employee's availability may be shared while its private reason remains restricted. The matched placebo replacements and separate artifact and reader scores provide a concrete way to investigate where exposure occurs. Furthermore, the fixed-transcript control examine the persistence of information in assistant replies after the original user fact is removed. Theses controls strengthen the interpretation that source information is being carried into shared artifacts, while making the limits of a simple leave-onefact-out intervention explicit.

Second, the paper makes substantial efforts to examine its measurement instruments. It includes calibration, alternative judges, model-assisted adjudication, independent human reannotation, and targeted examination of disputed labels. The statistical analysis also uses the persona as the clustering unit, preserving the paired comparison instead of treating repeated generations or reader questions as independent samples. The manuscript reports schema-dependent judge errors, distinguishes pooled extension results from independent replication, and retains missing outcomes separately from refusals or successful privacy protection. Although these checks reveal unresolved measurement problems, reporting them directly makes the evidence considerably easier to access.

Third, the authorized-sharing experiment evaluates actual task completion and downstream artifact resume. It checks publication to the intended audience, scheduling correctness, persistence after a session reset, and recovery by a separate reader receiving only a byte-identical digest copy. The inclusion of native audience isolation is also valuable. It establishes that useful sharing is already possible under an existing configuration. The paper appropriately reports that the confirmed-release configuration is not significantly superior to this baseline. The task and joint-success counts in Table 1 are also reproducible by aggregating the applied per-trajectory analysis record.

Critical weaknesses

However, there are concerns too. The major concerns are: (1) condition-dependent uncertainty in the central disclosure measurement, (2) implementation problems in the external-benchmark comparisons, (3) unresolved interpretation of the template effect, and (4) incomplete availability of evidence needed for independent reproduction.

First, the estimated profile-digest difference changes substantially with the labeling process. On the pools48-persona matrix, the model-assisted adjudicated contrast is 0.260, while the returned independent-human labels give 0.094. More importantly, the extension-only independent-human estimate is 0.055 with a confidence interval of [-0.008, 0.117], so the new cohort does not independently establish the contrast under that instrument. The subsequent targeted review raises the pooled estimate to approximately 0.255, but 174 of the 181 reviewed items were returned-positive by construction. This review can correct identified mistakes, yet it cannot estimate residual annotation error across a representative set of positive and negative artifacts. The disagreement is also closely related to the paper’s substantive question: annotators sometimes treat availability consequences as disclosure of the omitted private reason. Thus, uncertainty concerns the definition and measurement of the central outcome, rather than only ordinary sampling variation. The authors acknowledge this issue, but the reported persona-bootstrap intervals do not capture it. A representative independent annotation study with an explicit distinction between availability, private reason, and identifying detail would provide stronger support for the effect magnitude.

Second, the supplied external-benchmark implementations contain two concrete problems that weaken the supporting comparisons. In the Agent Leak runner, the model_inferred_tag condition replaces the ground-truth allowed field set with the tagger’s predicted_shared set before evaluating leakage. Consequently, a genuinely forbidden field incorrectly classified as shared is admitted into the context and then exempted from the leakage metric. An isolated check of the supplied scoring function confirms that the same explicit disclosure is counted under the true authorization set and ignored under an erroneous predicted set. The reported zero leakage therefore does not establish zero forbidden-field exposure under a common policy across conditions. Although the manuscript cautions that zero hits with 0.455 tag accuracy are not reliable enforcement, this does not resolve the change in the evaluation criterion itself. This issue directly affects the inferred-tag comparator; the original Trellis result and the other AgentLeak conditions require separate assessment.

Additionally, the supplied CI-Work reconstruction generates only three distinct writer inputs repeated across five direction labels. With the reported setting of three cases per direction, scenario selection uses the same first three base cases for every direction, and the writer prompt does not include the direction variable. The fifteen records therefore contain three unique prompt pairs, each repeated five times; the direction changes in the evaluator’s outline. This does not substantiate the claimed coverage of fifteen scenarios spanning five writer-side information flows. Since the historical per-item prompts are absent from the supplied package, the authors should clarify whether the reported execution used this exact construction. The effective scenario count and the interpretation of the external validation need correction.

Third, the main template comparison remains difficult to translate into a general principle for safer memory design. The profile and digest templates jointly change the organizing goal, field titles, and number of person-oriented slots. All primary borderline facts concern availability or scheduling, which directly matches some profile fields. The supplementary factorial experiments are useful and should be credited, but they do not establish a stable explanation across judges. At uniform capacity, the main judge does not detect either factor’s main effect, while some alternative-judge effects differ substantially. The apparent upper-range capacity effect also disappears under the alternative judges. Furthermore, the profile-digest ordering is not uniform in the generic-reply and marker controls. The evidence therefore supports a descriptive difference between particular templates more convincingly than a transferable rule about person-centered and event-centered memory. Broader evaluation with non-scheduling facts and semantically matched fields would strengthen the contribution beyond the observed template pair.

Fourth the supplementary package supports verification of several reported aggregates but does not provide all the material needed to independently audit the runtime and external comparisons. Its README explicitly states that approximately 10GB of raw runtime trajectories, external per-item records, and third-party repository snapshots are omitted. This is difficult to reconcile with the manuscript’s unqualified statements that raw responses and request/response bindings accompany the release. The supplied runtime records allow the Table 1 counts to be reconstructed, but they do not permit every privacy judgement, outbound publication, or reader-boundary check to be regenerated from the original evidence. The omission is particularly consequential for the AgentLeak issue, because the corrected leakage rate cannot be established from aggregate summaries. A compact export containing the relevant inputs, outputs, permissions, evidence quotes, and boundary records would substantially improve reproducibility without requiring the complete runtime archive.

The minor concerns are: (1) the restricted semantic scope of the runtime privacy result, and (3) the need to articulate the incremental contribution more directly.

First the runtime experiment covers eight fixed scheduling scenarios, two versions of one model family, and an ideal owner who checks a finite set of slot proposals. Its strongest privacy contrast concerns names declared private by the synthetic protocol. Under the frozen labels, private-reason disclosure is already 0/26 in Shared and 1/30 in Session. The later targeted review increases these counts, but mainly through coarse appointment or commitment descriptions. Moreover, some private reasons are summarized away before the downstream sharing tasks. These observations support a useful configuration audit, but provide limited evidence about protecting detailed sensitive reasons across more complex collaboration tasks. Independent human assessment and verification that the protected reason remains in private state would make this interpretation stronger.

Second the paper correctly acknowledges substantial overlap with prior work. PiSAs already evaluates cross-user information spillage through shared memory and other agent interfaces, while Fides studies deterministic enforcement of information-flow policies. The additional value here lies in the matched source interventions, template comparison, measurement analysis, and concrete runtime audit. The main text would benefit from explaining more directly with diagnostic conclusion or configuration decisions these additions enable beyond existing evaluations, particularly given the unresolved mechanism behind the template effect.

Questions for authors

The paper studies an important privacy boundary in shared agent systems and provides a useful combination of source interventions, artifact-level measurements, and executable authorized-sharing experiments. The distinction between source access and release authorization is clear, and the independent digest reader makes the runtime evaluation more informative. The authors also deserve credit for reporting annotation disagreements, incomplete evidence, and the absence of a demonstrated advantage over native isolation.

However, the contribution depends heavily on the reliability and reproducibility of its empirical audit. The central comparative effect remains sensitive to condition-dependent annotation, and the supplementary code reveals problems in two external comparisons that require correction or clarification. Missing raw evidence further limits independent assessment of those results. These concerns leave the current submission marginally below the acceptance threshold. A corrected common-policy evaluation, an accurate account of the CI-Work scenario diversity, and a sufficient evidence release would materially strengthen the case for acceptance.

Here are several questions for authors: First, can the authors re-score the AgentLeak inferred-tag condition against the immutable scenario authorization set, keeping the evaluation policy identical across all conditions? In particular, how do the zero-leakage result and the per-channel comparisons change when incorrectly shared fields remain forbidden for evaluation?

Second, can the authors clarify whether the reported CI-Work run used the supplied scenario generator? That implementation produces three unique writer inputs repeated across five direction labels. Please provide the actual prompts and explain which results support coverage of five distinct information-flow directions.

Third, how should readers reconcile the pooled schema contrasts of 0.260 under model-assisted adjudication, 0.094 under independent human reannotation, and approximately 0.255 after targeted review? Can the authors provide representative independent adjudication or an additional sensitivity analysis that explicitly separates availability disclosure from disclosure of its private reason?

Fourth, can the authors provide a compact, complete evidence bundle for independently reproducing the external comparisons and runtime privacy checks, including per-case outputs, true permissions, annotation evidence, and the relevant publication and reader-boundary records? This would also clarify the difference between the manuscript’s reproducibility claims and the omissions documented in the supplementary README.

2 comments

Thank you for the careful checks. Both external-benchmark criticisms are correct.

Q1. AgentLeak. The old 0.000 was wrong: predicted_shared both admitted fields and replaced effective_allowed during scoring, so a forbidden field marked as shared was admitted and excluded from the metric. We re-scored all 1,000 scenarios against immutable allowed_set.fields on existing outputs; the other three conditions are unchanged item by item:

ConditionsinglemultiC1C2C5internal-only
unbound0.4020.8450.6240.8060.6950.183
prompt policy0.2620.4410.3140.3700.3210.058
trusted scope0.0030.0230.0160.0190.0050.004
inferred tag (corrected)0.3640.6860.5510.6530.5730.102

Here single means leakage in a single channel and multi means leakage across multiple channels. C1 is the final output, C2 is inter-agent messages, and C5 is shared memory. inferred_tag multi goes from 0/1000 to 686/1000. In 956/1000 scenarios, the tagger admitted at least one forbidden field; tag accuracy stays 0.455. Only this comparator changes, under the same lexical oracle. The 0.023 trusted_scope residual comes from that matcher. Its mechanism is unknown, and we do not call it semantic leakage.

Q2. CI-Work. You are right. The generator picked the same first three base cases for every direction, and build_writer did not use direction; it appeared only in the evaluator outline. Each condition tested three different writer inputs, each repeated five times. The official15b counts (10/15 neutral, 0/15 per defense) are repeated-run counts, not 15 scenarios or five flows. Historical per-item prompts were not saved and cannot be recovered verbatim. The revised text gives the effective scenario count as three.

Q3. The different contrasts. These numbers come from different label processes and are not interchangeable:

  • 0.094 (48 personas) and 0.055 (extension 32, 95% CI [-0.008, 0.117]): returned independent-human labels. The extension does not independently establish the contrast.
  • 0.260: the model-assisted adjudicated reference on the pooled matrix.
  • 0.255 (pooled; 0.266 extension): a targeted review of 181 disagreement-flagged items. Of these, 174/181 were returned-positive by construction, so it cannot estimate residual error across a representative set of positive and negative artifacts.

Some artifacts kept only the scheduling consequence, yet were labeled as disclosing the omitted private reason. So the disagreement is also about what counts as disclosure: availability, private reason, or identifying detail. Persona-bootstrap intervals do not cover this uncertainty. We have not run a new representative human study separating those outcomes, so the revised paper treats effect size as label-definition dependent.

PVR means purpose violation rate, measured here as P(Au≥2)P(A_u \ge 2): the fraction of artifacts with paraphrase-level exposure of the private reason or fact core. It does not count every violation of the strict no-release prompt, which also withholds availability; we did not recompute Au≥1A_u \ge 1.

Q4. Evidence and what is included in this revision. We upload only the revised PDF. The original supplementary ZIP stays unchanged, and the compact archive described below is not included.

Our submission overstated the supplement's completeness. It omitted about 10 GB of runtime trajectories, external records, and third-party snapshots. We narrowed the Reproducibility Statement and fixed a typo reporting Session private-name hits as 7 instead of 9, matching Session's [9,11]/32 bounds; the two unjudged runs are Shared.

Offline, we built a 50.5 MB archive of all 152 runtime runs (146 completed; incomplete runs are not privacy negatives), binding all 160 annotation calls and 146 digest-transfer hashes, plus 4,000 AgentLeak common-policy rows, MuPPET records, and reconstructed CI-Work inputs. Blocking outside reads reproduced all aggregates and boundaries. This checks saved automated labels, not semantic truth or independent human agreement. CI-Work prompts are unrecoverable, and raw traces need anonymization and license review before release.

Other points. Profile–Digest describes only this template pair, not a general rule. LOFO supports the private source's contribution to disclosure, not an internal stage. Confirmed Release (62 simulated-owner commands, not human approvals) shows no security gain over Native Isolation. Runtime covers eight scheduling tasks and mainly protocol-declared names (frozen reason leaks: 0/26 Shared, 1/30 Session); early summarization can remove reasons. We credit PiSAs for spillage and limit our increment to source counterfactuals, artifact/reader checks, annotation analysis, and runtime configuration evidence.

Thank you for your response. I will maintain my current rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.