acceptodds
Withdrawn from ICLR 2027

Mind the Boundary: Data Leakage in Native-Store LLM Browser Extensions

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

LLM browser extensions legitimately send some user content to remote models, making outbound traffic alone an inadequate privacy signal. We present a four-browser measurement study that first identifies LLM-related candidates, then uses package analysis to characterize their permissions and model-facing data surfaces, and finally executes downloadable extensions with synthetic markers to observe first-hop transmission. Among 6874 successfully analyzed candidate packages, The risks identified through static analysis, such as unnecessary browser permissions and AI-related content collection are widespread. Runtime evidence is substantially narrower: most detected transmissions carry explicit user input, while a small subset also carries context outside our predeclared profile-and-action boundary. A controlled context experiment further shows that richer and attacker-influenced pages can induce additional out-of-bound transmissions. Our results have led to the separation of static risks and risks under actual operation, and have revealed the proportion and correlation between the two in the real world.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Browser Extension Security, Web Security, Data Leakage

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5kKaOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper conducts an empirical study of LLM-based brower extensions. The goal is to identify what use (sending user data to a server) is for legitimate (extension functionality related) uses and what is not (e.g., a privacy violation). For Chrome and Firefox, several thousand extensions are used, along with a few hundred for Edge and about a dozen for Opera.

Critical strengths

The appear addresses an interesting question of growing importance.

Critical weaknesses

The paper is written so poorly that is it genuinely hard to understand or follow. Separate from the many typos, grammatical errors, and massively repetitive prose, too much is entirely unclear. For example: a. What "following text" is collectively referred to as LLM extension? b. In table 1 (which, as best I can tell, is presented but never referenced), what do the last three columns mean? c. What does it mean that "Chrome combines 132,356 unique manifests ... with 33 current-store queries? d. Why were ~800 packages not successfully scored?

At a higher level, though, the actually process that authors went through to obtain the empirical results presented is too opaque.

Questions for authors

The paper requires significant revision and proof reading before being considered for acceptance. The questions I have are too vague (e.g., what did you actually do?) to yield fruitful discourse here.

Reviewer_m1TvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a cross-store study of privacy risks in LLM browser extensions. It introduces a profile-aware legitimate-input boundary to distinguish task-relevant data transmission from over-collection, combining rule-based static analysis with dynamic testing using synthetic markers. The authors analyze 6,874 extension packages and collect 5,993 usable runtime traces, identifying six extensions with seven baseline boundary violations. Controlled experiments further show that complex and attacker-influenced page contexts can trigger additional violations.

Critical strengths
  1. Evaluates data transmission against expected functionality rather than treating all outbound LLM traffic as suspicious.
  2. Analyzes thousands of extensions across four stores and separates static risk signals from observed runtime behavior.
  3. Investigates how page complexity and attacker-influenced text affect data transmission while avoiding unsupported prevalence claims.
Critical weaknesses
  1. Profiles are inferred from metadata and manifests without independent ground-truth validation. A single profile may misclassify multifunction extensions.
  2. Generic scoring outperforms the full profile-aware score for remote-other behavior (AUROC 0.7059 vs. 0.6520) and boundary-positive prediction (0.7603 for generic scoring). The practical advantage of the proposed scoring remains unclear.
  3. First-hop-only observation, untriggered features, and potential failures to detect transformed content may lead to missed violations. Only six extensions show baseline violations, limiting the strength of prevalence-related conclusions.
Questions for authors
  1. How accurate are the assigned profiles and legitimate-input boundaries? Can a manually validated subset or multi-profile analysis quantify misclassification?
  2. Given the AUROC results, what measurable benefit does profile-aware scoring provide over generic checks at a fixed review budget?
  3. How does sentinel detection handle encoded, compressed, or transformed payloads? Have you measured its sensitivity?
  4. How might untriggered features and unobserved relay or backend flows affect the findings? Can a controlled subsample help estimate the extent of missed violations?
Reviewer_QmDQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper collects 7,688 candidate extensions across Chrome, Firefox, Edge, and Opera, successfully analyze 6,874 packages, and obtains 5,993 usable baseline runtime traces. A deterministic classifier assigns one of eight functional profiles, and 43 static rules describe generic capabilities, AI-related surfaces, and profile-relative overreach. The dynamic harness exercises extensions using synthetic markers and inspects first-hop requests.

In the baseline, 79 extensions produce remote transmissions containing markers. Six extensions produce seven events classified as exceeding their profile-and-action boundary. A separate experiment tests a baseline-responsive cohort across synthetic page families and complexity levels, finding more boundary-positive outcomes on high-complexity pages.

Critical strengths
  1. The target problem on security and privacy of LLM browser extension is useful and timely.

  2. The study covers a substantial number of packages and several native stores.

  3. The paper distinguishes first-hop destination from payload contents and avoids treating an opaque relay or a known provider as automatic evidence of misuse.

  4. The interaction ablation is informative.

Critical weaknesses
  1. The task boundaries lack independent validation, making it difficult to determine whether the reported positives represent unnecessary data collection.

  2. The context experiment changes page complexity and attack content together, so it cannot isolate attacker influence from ordinary broad context extraction.

  3. The practical value of profile-aware scoring remains unclear. Generic scores rank observed positives better, while limited runtime coverage makes negative outcomes difficult to interpret.

Major Concerns

  1. The boundary definition is not validated well enough to support the central labels.

A deterministic profile assignment is reproducible, but it is not necessarily correct. The ordered regular-expression classifier assigns a single profile and defaults unmatched extensions to general copilot. The paper does not evaluate these assignments against independent human judgments, nor does it provide a complete profile-by-action-by-marker policy specifying which content is allowed under each exercised workflow.

The checks in Appendix A do not substitute for this validation. Agreement between metadata-only and metadata-plus-manifest classification measures stability of two related heuristics. Similar profile distributions across stores likewise do not establish that individual classifications or permitted-input boundaries are accurate.

  1. RQ3 does not isolate attacker influence from broader input collection.

High-complexity pages add hidden/contextual fields and attack-bait, while low-complexity pages contain task-essential surfaces. Comparing the same extension across these pages controls extension identity, but several treatment components still change together. More importantly, some markers that can make a request boundary-positive are introduced only in the high condition. An extension using the same broad extraction routine could therefore become newly positive simply because the page now contains an additional disallowed field. The attack-bait result has a related ambiguity. Seeing instruction-like page text in an outbound request establishes transmission, not that the extension or its model followed those instructions. The discussion recognizes overbroad context assembly as a possible mechanism, but the causal framing around attacker influence remains stronger than the experiment supports.

  1. Dynamic coverage and the responsive-cohort selection are too weakly characterized.

Only 79 of 5,993 extensions with usable traces produce a remote marker-bearing event in the baseline. This is not evidence that the rest do nothing, as the paper acknowledges. However, the study does not independently measure how often its scripted actions actually reach the advertised feature. A usable trace may show that instrumentation worked while the extension remained blocked on onboarding, authentication, an unsupported page, or an unrecognized interaction.

The interaction ablation measures recovery relative to already observed positives. It does not estimate how many positives remain undetected outside that cohort. A stratified manual test of apparently negative extensions would make the negative outcomes much more interpretable.

RQ3 also needs its cohort size, exact enrichment criteria, per-browser/profile composition, and matched denominators for every condition. Percentages such as 25.00% for summarizers are hard to assess without counts. Repeated runs, uncertainty intervals, and analysis clustered by extension are needed because the same extension appears across page families and context levels. I agree that these results should not estimate Web prevalence, but even the within-cohort comparisons need a clearer experimental population.

  1. The static ablation does not demonstrate an improvement in triage.

Adding nonnegative rule weights raises the high/critical count from 1,303 to 2,492 and then 2,898. At a fixed score threshold, an increase is expected by construction. It does not show that generic analysis misses genuine privacy problems, or that the additional 1,595 packages are better review targets.

Appendix C.1 is more informative: generic scores achieve AUROC 0.7059 for remote-other outcomes, compared with 0.6520 for the full profile-aware score. Generic scores also perform best for boundary positives, with AUROC 0.7603. This does not invalidate using profiles to interpret a payload, but it leaves the proposed scoring system's practical benefit unestablished.

For the proposed review-queue use case, precision or yield at a fixed review budget would be more useful, alongside precision-recall curves and uncertainty. With only six baseline boundary-positive extensions, the stability of any ranking comparison is itself a concern. The paper should also avoid using "calibration" for a result that measures ranking discrimination rather than probability calibration.

  1. The methods and case evidence are not sufficiently specified for reproduction.

Appendix E provides rule names, weights, and counts, which is useful. It does not provide the exact detection patterns or establish whether a source and a model-facing sink merely coexist in a package or are connected by a feasible data flow. This distinction is essential for interpreting claims such as form-input-to-LLM coupling, especially in bundled libraries or dead code. The exact severity cutoffs are also not given in the supplied manuscript, despite the contribution statement saying they are exposed.

Technical and Presentation Issues

  1. Define "remote-other-positive" before using it in Section 4.2. It is not one of the labels defined in Section 3.3. Also reconcile the 77 remote-other positives with the 48 high/critical and 26 non-high/critical positives, which sum to 74. If this is a restricted static/dynamic join, give its denominator and exclusions.

  2. Report precision of the candidate-selection heuristic on a human-checked sample. The 5,151 packages with an endpoint or an LLM-specific code shape provide stronger evidence than metadata alone, but the remaining candidates still enter several LLM-extension analyses. Show whether the main results change on the stronger-evidence subset.

Questions for authors
  1. Can you provide the complete profile/action boundary policy and independent human validation of the profile assignments and all six baseline-positive decisions?

  2. What happens when attack-bait is replaced with matched neutral text while the DOM layout and sensitive-marker opportunities remain identical? Does the extension transmit additional sensitive content, or simply forward the visible instruction text?

  3. What are the exact RQ3 cohort and matched sample sizes, and how often does manual testing recover activity from extensions labeled negative by the automated harness?

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

Related papers

Discussion (0)

Sign in to comment.