{"id":"3110605a-e1e6-4241-93ed-849b9f8706b4","arxiv_id":"2608.13390","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"More than half of 278 Telegram Mini Apps sent data to undisclosed third parties, mostly at page load, and none displayed a consent mechanism at launch.","lead":"Researchers built a tool that intercepts network traffic from 278 Telegram Mini Apps and compared the services those apps actually contacted with their privacy policies. Over half contacted at least one third party that the policy did not mention, and none showed a consent prompt at launch.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 59.4% violation rate hinges on the interpretive mapping of Telegram's generic Standard Bot Privacy Policy; if sharing to domains not named is not actually prohibited, the headline overstates non-compliance.","rationale":"The reader's weakest_assumption identifies exactly the same hinge: the operationalization of the Standard Bot Privacy Policy in Section 4/Table 1. I agree that this is the load-bearing point, and the proposed severity is CONDITIONAL. I considered whether the consent-zero finding could be a stronger concern, but the paper itself already scopes it carefully in Section 9.1 and frames it as absence of observable mechanisms rather than a legal conclusion, so the abstract's wording is the only overreach there. The policy-mapping concern is more central because it determines whether any violation exists at all for the majority of apps. The paper does provide independent support in the form of reproducible tooling and data release (Section Data availability), and the conservative manual-review criterion for custom policies genuinely biases toward under-reporting. However, that conservative criterion is applied only to the 59 custom-policy apps; for the 219 Standard-policy apps the authors apply their own interpretation without an equivalent conservative re-read. That asymmetry is the concrete soft spot. I therefore agree with the reader's CONDITIONAL verdict: the direction is plausible and worth publishing, but the central rate should be presented as dependent on an interpretive mapping and ideally validated. I would not REJECT because the concern is about the strength of a headline rate, not an internal contradiction.","tokens_in":22177,"tokens_out":1683,"duration_ms":110149,"concrete_test":"Take the 100 most frequently appearing undeclared domains from the withheld per-app annotations and classify each by declared purpose (analytics, ad network, hosting/CDN, crash reporting, etc.). Re-label a Standard-policy app as non-violating when every undeclared contact is to a domain whose provider purpose falls inside a category that the Standard Bot Privacy Policy could plausibly permit under the necessity/purpose-limitation and 'not transferred unless authorized' clauses (e.g., service-operational analytics or content delivery). If the re-labelled violation rate drops materially below 59.4%, the headline number is an artifact of a strict interpretive mapping; if it stays above, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that 165/278 apps (59.4%) commit a privacy-policy violation by contacting a third-party domain not disclosed in the applicable policy. For the 219 apps governed by Telegram's Standard Bot Privacy Policy, this conclusion rests entirely on the authors' interpretation in Section 4 that the default policy's prohibition on disclosing private user information to third parties without explicit user authorization, combined with its requirement of necessity and purpose limitation, makes any outbound request carrying user-related data to a domain not listed in the policy a violation. But the policy is generic and never enumerates recipient domains, and a purpose-limited analytics or infrastructure provider could plausibly fall within 'processing only insofar as necessary for the designated features' or within categories that the policy allows without naming specific domains. The authors acknowledge this interpretive burden in Section 9.2 for custom policies but do not validate their reading of the Standard policy against Telegram or a competent authority. If a domain is covered by an allowed category (e.g., hosting, analytics necessary for functionality) rather than prohibited, the violation count could shrink substantially. The custom-policy finding that recipients described by category count as disclosed when a policy could 'plausibly be read as covering it' highlights the asymmetry: domains contacted under the Standard policy are not treated with the same generous reading.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents TeleGapper, a black-box dynamic analysis framework for Telegram Mini Apps. The framework takes only a bot name, launches the Mini App in a rooted Android Telegram client, intercepts HTTPS traffic through a MITM proxy, separates traffic at page load from traffic during randomized user-driven exploration, and compares the contacted domains with the applicable privacy policy. The authors collected 991 catalogue entries from tApps Center, successfully analyzed 278 working Mini Apps, and report that 59.4% contact at least one third-party domain not disclosed by the applicable privacy policy, that 78.8% rely on Telegram's Standard Bot Privacy Policy, and that none of the 278 apps displayed a consent or opt-out mechanism at launch. The paper also reports that violations are concentrated in a small set of analytics and advertising services, that 85.5% of violating apps contact undisclosed third parties already at page load, and that device and Telegram profile information are the most commonly transmitted data types. The authors position this as the first systematic empirical study of privacy-policy compliance in Telegram Mini Apps.","tokens_in":22440,"tokens_out":6003,"duration_ms":57887,"significance":"If the results hold, this paper makes a valuable contribution: it is the first black-box dynamic analysis framework for Telegram Mini Apps, it is applied to a reasonably large sample, and it connects runtime network behavior to the applicable privacy policy, addressing a largely unexplored ecosystem. The authors should be credited for releasing the scraper and the 991-entry dataset, for documenting the pipeline in detail, and for explicitly acknowledging several threats to validity. The classification of third-party services via AppBrain and the reported statistical tests for RQ2 are also strengths. The main significance, however, rests on the interpretation that any contact with a domain not named in Telegram's generic Standard Bot Privacy Policy constitutes an undisclosed third-party communication; if that mapping is too broad, the headline violation rates would overstate non-compliance. The absence-of-consent finding is clearly scoped in Section 9.1 and is a useful empirical observation independent of the policy-interpretation issue.","major_comments":[{"comment":"The headline result (59.4%, Table 2) is defined in Section 5.4 as 'transmission of user data to a third-party domain that is not disclosed by the privacy policy.' For the 219 apps governed by the Standard Bot Privacy Policy, 'not disclosed' is operationalized in Section 4 as 'not identified in the applicable privacy policy,' and the policy never enumerates recipient domains. Consequently, any third-party contact carrying user-related data is automatically classified as undisclosed. This is an interpretive mapping, not a direct consequence of the policy text: a purpose-limited analytics or infrastructure provider could plausibly fall within the policy's allowance of processing 'only insofar as necessary for the designated features' or within an 'otherwise permitted' category. The authors do not validate their reading against Telegram or a supervisory authority, and Section 9.2 acknowledges interpretive discretion only for custom policies, not for the standard policy. If this mapping overstates what the policy prohibits, the RQ1 and RQ2 conclusions would overstate non-compliance. Please provide the full relevant policy language, or a legal-analysis/validation step, or a sensitivity analysis under more restrictive readings (for example, counting only contacts whose service category has no plausible functional justification).","section":"Section 4, Table 1, and Sections 5.4 and 7.1"},{"comment":"The EXCLUDED_HOSTS set filters out hosts including fonts.googleapis.com and play.googleapis.com before third-party inspection. These hosts can be contacted by Mini App code itself (for example, a WebView page loading Google Fonts) rather than only by the Android operating system. The set was constructed by manually inspecting traffic from 100 randomly sampled apps, but the paper does not report how many of the 278 analyzed apps were affected by this filter or whether the excluded hosts appeared in application-level requests. Section 9.2 correctly notes the filter as a false-negative source, but the filter can also change per-app violation classification and therefore the denominator of Table 2. Please quantify the sensitivity to this exclusion, for instance by re-running the analysis with these hosts included and reporting the change in the number of violating apps and the violation rate.","section":"Section 6.5 and Section 9.2"},{"comment":"The review of custom privacy policies relies on a single reviewer applying a 'plausibly can be read as covering it' criterion. No inter-rater reliability measure is reported, and the per-app annotations are not released, as stated in the Data and code availability section. Because 33 of the 59 custom-policy apps are classified as violating at least partly on the basis of this judgment, the custom-policy result is not independently verifiable. The stated criterion is conservative and biases toward under-reporting, which is reassuring, but the magnitude of the interpretative discretion is not quantified. Please either release the annotations under a redaction scheme that avoids the personal-data concern they describe, or provide a second annotator and report agreement; if neither is feasible, the custom-policy violation counts should be reported with a confidence range reflecting annotator disagreement.","section":"Sections 5.5 and 9.2"}],"minor_comments":[{"comment":"The 20-second initialization window and the 240-second exploration budget are presented as empirical choices without sensitivity analysis; a short paragraph reporting how results vary with, for example, a 10-second or 40-second initialization window would strengthen the timing claims in Section 7.3.","section":"Section 6.2"},{"comment":"The regular-expression extractor for initData requires values of at least 8 characters; very short identifiers or numeric-only user IDs could be missed, and the false-negative rate of the extractor suite is not evaluated. A brief discussion of this limitation would be useful.","section":"Section 6.4"},{"comment":"The RQ2 summary states that exposing a custom policy 'does not correspond to a statistically distinguishable improvement,' which is appropriate, but the abstract's phrasing 'custom policies show no statistically significant improvement' could be read as a claim of no effect; the power analysis in Section 7.2 itself shows the test is underpowered for the observed effect size, so the wording should be aligned with 'no significant difference was detected.'","section":"Section 7.2"},{"comment":"The abstract states that 'none provides a consent or opt-out mechanism,' but Section 9.1 limits this finding to the launch screen and the sessions observed during analysis. The abstract should carry the same qualification, for example 'none presented an observable consent mechanism at launch in our executions.'","section":"Section 7.3 and Abstract"},{"comment":"Minor presentation issues: 'FindiMini.app' is spelled inconsistently; the dataset name 'DtApps' is typeset with a formatting artifact in Section 7; and table headers such as 'T otal' have inconsistent capitalization.","section":"Throughout"}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is the first systematic measurement I've seen of privacy-policy compliance in Telegram Mini Apps. The setup is a black-box dynamic analysis framework that launches each app, intercepts HTTPS traffic via a MITM proxy, separates page-load from interaction traffic, and compares contacted domains against whatever policy applies. They ran it on 278 working apps sampled from tApps Center. The headline result — 59.4% contact at least one undisclosed third party — is new and probably in the right direction, but the exact number depends on an interpretive call about Telegram's default policy, and that is the one place I'd push back.\n\nWhat they do well: the methodology is coherent and honestly documented. The load/interaction split is a good design choice and produces a robust observation: 85.5% of violating apps already contact undisclosed third parties at page load, before any user interaction is possible. The absence of any observable consent or opt-out mechanism across all 278 apps is also striking, and the body is appropriately careful to scope it to the launch screen and observed sessions. The comparison between custom-policy and default-policy apps uses standard tests plus a power analysis, which is the right level of rigor. The related-work coverage is solid; the authors correctly distinguish their black-box runtime approach from the static taint-analysis methods used for WeChat mini-apps, and the self-citation to their PolicyGapper paper is on topic.\n\nThe soft spots. First, the load-bearing one: for the 219 apps under Telegram's Standard Bot Privacy Policy, a violation is defined as any user-data-bearing request to a domain not named in the policy. But the policy never names domains; it prohibits transferring private user info without user authorization or unless 'otherwise permitted by the policy,' and it permits processing 'necessary for the designated features.' A purpose-limited analytics or hosting provider could plausibly fall under that allowance. The authors flag the interpretive burden for custom policies, where they deliberately read recipients generously, but they do not apply the same generous reading to the default policy. If a contacted domain is covered by an allowed category, the 59.4% figure overstates non-compliance. I think this is a moderate concern, not fatal: the most frequent contacted domains are ad networks and tracking services (Google Tag Manager, Yandex Metrica, Adsgram), which are hard to justify as 'necessary' for most of these apps.\n\nSecond, the abstract overstates the consent finding: 'none provides a consent or opt-out mechanism' goes beyond the evidence, which is that none appeared on the launch screen in their runs. The body gets this right.\n\nThird, raw traces and per-app annotations are not released. The authors explain that traces contain their own Telegram account data, which is fair, but it means the central measurements cannot be independently verified. The released scraper and catalog dataset are useful; per-app labels would be more so.\n\nNet: this is a legitimate first measurement of an important ecosystem. It deserves peer review. The authors should engage the interpretive issue head-on — justify their reading of the default policy or add a sensitivity analysis under a more permissive reading — and should add confidence intervals to the headline rates. The consent finding is the most defensible contribution.","headline":"First real measurement of Telegram Mini App privacy-policy compliance, with an interpretation-dependent violation rate and a strong consent finding.","tokens_in":22923,"tokens_out":4411,"duration_ms":36262},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Telegram Mini Apps routinely send user data to third parties their own privacy policies never name, and the ecosystem offers users no consent screen to stop it.","keywords":["Telegram Mini Apps","privacy policy compliance","dynamic analysis","black-box analysis","third-party tracking","network traffic interception","consent mechanism","GDPR transparency"],"falsifier":"Ask Telegram (or a data protection authority) for an authoritative interpretation of which recipient categories the Standard Bot Privacy Policy covers; if services such as Yandex Metrica, Google Tag Manager, or Adsgram fall inside the policy's intended scope, the headline violation rate would be a measurement artifact rather than evidence of non-compliance. An alternative check is to contact the developers of the 59 custom-policy apps and ask whether their policy was meant to cover the undeclared domains; if most say yes, the manual annotation is too strict.","tokens_in":21972,"feed_emoji":"📡","tokens_out":10328,"duration_ms":288083,"temperature":0.7,"pith_summary":"Telegram Mini Apps are ordinary web pages that run inside the Telegram client and receive a user's IP address and Telegram profile data automatically at launch, before the user has acted. This paper asks whether the privacy policies governing those apps say where that data actually goes. Using a black-box tool that launches 278 working Mini Apps, intercepts their HTTPS traffic, and compares each contacted domain with the applicable privacy policy, the authors report that 59.4% of the apps contact at least one third party their policy does not name, that 78.8% rely on Telegram's generic default policy, and that not one app shows a consent mechanism at launch. The result matters because the privacy policy is effectively the only transparency instrument users have in this ecosystem, and it appears largely decoupled from actual data flows.","feed_headline":"59.4% of Telegram Mini Apps share data with undeclared third parties","feed_subtitle":"No consent banners and page-load tracking appear in most of 278 apps checked against their policies.","key_machinery":"The central mechanism is TeleGapper, a black-box dynamic analysis pipeline that takes only a bot name as input, launches the Mini App inside the Telegram client on a rooted device, intercepts outbound HTTPS traffic through a man-in-the-middle proxy, separates traffic produced at page load from traffic produced by randomized user-driven exploration, and filters out platform infrastructure hosts. On the policy side, the authors derive five violation classes from Telegram's Standard Bot Privacy Policy: undisclosed third-party communication, third-party transmission of user data, potential over-collection, pre-interaction third-party communication, and tracking or advertising communication. The verdict for each app comes from comparing the observed third-party domains with the recipients named in the applicable policy, where a custom policy counts as covering a recipient whenever it can plausibly be read to do so, and the default policy counts as covering only what it explicitly names.","core_discovery":"The paper's central claim is that privacy-policy compliance is the exception in the Telegram Mini App ecosystem. Across 278 randomly sampled working apps observed over five runs each, 59.4% (165) transmit user-related data to at least one third party not disclosed by the applicable privacy policy, with a mean of 2.78 undeclared recipients per violating app and a maximum of 14. The violations are front-loaded: 85.5% of violating apps contact undisclosed third parties during page load, before any meaningful in-app interaction, and none of the 278 apps presents an observable cookie banner, consent dialog, or opt-out control. The data flowing to these recipients is not limited to technical parameters: 95.2% of violating apps transmit device information, 38.2% transmit user and profile information such as the Telegram user ID and initData payload, and one in three transmits both. A further 78.8% of apps rely on Telegram's Standard Bot Privacy Policy, and apps with custom policies violate at a statistically indistinguishable rate (60.3% vs 55.9%, p = 0.55).","pith_inferences":["The near-identical violation rates for custom and default policies suggest that publishing a privacy policy in this ecosystem is not a meaningful commitment to disclosed practices; a stronger test would correlate policy publication with other developer signals such as app age, update frequency, or responsiveness to disclosure requests.","Because the same small set of ad and analytics services recurs across unrelated apps, a platform-enforced allowlist of permitted recipient domains would likely eliminate most of the measured exposure; this is testable by simulating such an allowlist against the released dataset.","The black-box method transfers directly to other WebView-hosted mini-app ecosystems and to ordinary web applications with an implicitly authenticated session, so the transparency gap documented for Telegram probably is not unique to it."],"forward_implications":["If these rates generalize beyond the sampled catalogue, most Telegram Mini Apps cannot serve as transparent data processors: the privacy policy users are shown does not reveal the actual recipients of their data.","A platform-side runtime check modeled on TeleGapper would flag the majority of Mini Apps, since violations concentrate in a small set of recurring analytics and advertising services.","Because 85.5% of violating apps already phone home at page load and no consent mechanism is visible at launch, the default \"continued use constitutes acceptance\" model cannot plausibly ground these transfers.","Because Mini App content is served from developer-controlled URLs and can change without redistribution, any one-off compliance result is a point-in-time snapshot and cannot support a lasting claim of compliance."],"supporting_citations":[{"why":"Telegram's Standard Bot Privacy Policy supplies the normative text from which the paper derives its violation classes and the default regime that governs 78.8% of the sampled apps.","marker":"[30]"},{"why":"Telegram's Terms of Service for Mini Apps establishes which user data a Mini App receives automatically at launch, including IP address, profile metadata, and initData.","marker":"[33]"},{"why":"The AppBrain catalogue provides the externally maintained service categories used to label undeclared recipients as analytics, advertising, or tracking.","marker":"[3]"},{"why":"A prior annotation protocol for comparing app privacy policies against observed data flows is used for the manual review of the 59 custom policies.","marker":"[8]"},{"why":"GDPR Article 13 establishes the transparency obligation that makes undisclosed third-party recipients a legal concern.","marker":"[13]"},{"why":"ePrivacy Directive Article 5(3) establishes the prior-consent requirement that makes the absence of consent banners at page load relevant.","marker":"[6]"}],"fun_headline_variants":["TeleGapper: 59.4% of Telegram Mini Apps hide data sharing from users","No consent banners as 59.4% of Telegram Mini Apps contact unknown parties","78.8% of Mini Apps rely on generic policy while 59.4% leak data","Most Telegram Mini Apps violate privacy policies during first load","Telegram Mini Apps: 59.4% send data to undisclosed third parties"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the paper's own reading of Telegram's Standard Bot Privacy Policy as prohibiting any user-data transmission to domains the policy does not explicitly name; if Telegram or a supervisory authority interprets the policy's generic recipient language more broadly, the 59.4% violation rate would overstate actual non-compliance.","fun_headline_variants_meta":{"raw":{"variants":["TeleGapper: 59.4% of Telegram Mini Apps hide data sharing from users","No consent banners as 59.4% of Telegram Mini Apps contact unknown parties","78.8% of Mini Apps rely on generic policy while 59.4% leak data","Most Telegram Mini Apps violate privacy policies during first load","Telegram Mini Apps: 59.4% send data to undisclosed third parties"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000632,"raw_usage":{"total_tokens":2980,"prompt_tokens":1069,"completion_tokens":1911,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":1805}},"tokens_in":685,"tokens_out":1911,"duration_ms":12441,"temperature":1.0,"reasoning_tokens":1805,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:00:32.462380+00:00","model_set":{"reader":"deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.0-daily-deepseek-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Ask Telegram (or a data protection authority) for an authoritative interpretation of which recipient categories the Standard Bot Privacy Policy covers; if services such as Yandex Metrica, Google Tag Manager, or Adsgram fall inside the policy's intended scope, the headline violation rate would be a measurement artifact rather than evidence of non-compliance. An alternative check is to contact the developers of the 59 custom-policy apps and ask whether their policy was meant to cover the undeclared domains; if most say yes, the manual annotation is too strict.","supporting_citations":[{"cited_title":"Standard Bot Privacy Policy.https://telegram.org/ privacy-tpa, Accessed August 14, 2026","cited_arxiv_id":null,"evidence_quote":"Telegram's Standard Bot Privacy Policy supplies the normative text from which the paper derives its violation classes and the default regime that governs 78.8% of the sampled apps."},{"cited_title":"Terms of Service for Mini Apps.https://telegram.org/ tos/mini-apps, Accessed August 14, 2026","cited_arxiv_id":null,"evidence_quote":"Telegram's Terms of Service for Mini Apps establishes which user data a Mini App receives automatically at launch, including IP address, profile metadata, and initData."},{"cited_title":"AppBrain: Android Ad Network statistics and market share.https://www.appbrain.com/stats/libraries/ad-networks/, accessed August 14, 2026","cited_arxiv_id":null,"evidence_quote":"The AppBrain catalogue provides the externally maintained service categories used to label undeclared recipients as analytics, advertising, or tracking."},{"cited_title":"Policygapper: Automated detection of inconsistencies between google play data safety sections and privacy policies using llms, 2026","cited_arxiv_id":null,"evidence_quote":"A prior annotation protocol for comparing app privacy policies against observed data flows is used for the manual review of the 59 custom policies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GDPR Article 13 establishes the transparency obligation that makes undisclosed third-party recipients a legal concern."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ePrivacy Directive Article 5(3) establishes the prior-consent requirement that makes the absence of consent banners at page load relevant."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}