{"id":"3abcadb0-9a4f-4cb6-9f7f-3ced589f213f","arxiv_id":"2611.13965","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A qualitative comparison shows WeChat Mini Programs run in a platform-managed runtime while Telegram MiniApps are web apps in a WebView, and in both ecosystems developers remain responsible for backend authorization.","lead":"This paper compares how WeChat and Telegram host small third-party apps, focusing on where security enforcement actually lives. It finds that WeChat runs MiniApps in a dedicated platform runtime while Telegram loads them as ordinary web pages inside a WebView, with both platforms leaving backend authorization to developers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Telegram source inspection is a single-revision snapshot; the paper's central architectural claims about Telegram's permission mediation (E3–E5) may not generalize across client versions or platforms.","rationale":"The reader's weakest_assumption correctly identifies that the Telegram-specific findings rest on a single manual reading of one commit. I agree that this is the most load-bearing concern. The paper is transparent about this limitation, and the verdict ACCEPT with HIGH confidence is reasonable given the paper's stated scope as a qualitative systematization. However, I want to emphasize that the concern is not merely an acknowledged scope limit; it is structurally load-bearing because the paper's most distinctive and actionable claims (E3–E5, Worked Example 2) depend on the correctness and stability of that single-revision reading. If the code changes, the modeling requirements lose their concrete grounding. The concrete test I propose would settle whether this concern actually lands: if the patterns are stable across versions, the concern is mitigated; if they are not, the paper's central architectural claims about Telegram are version-specific and the derived requirements would need qualification. I recommend UNCHANGED because the paper already states the limitation and does not overclaim generality. The reader's verdict is appropriate; my analysis reinforces the same concern rather than identifying a new one.","tokens_in":16983,"tokens_out":699,"duration_ms":8571,"concrete_test":"Inspect the same code areas (BotWebViewContainer, ArticleViewer, bridge registration, permission handlers, multi-window navigation) in the current HEAD of the Telegram Android repository and in at least two earlier tagged releases (e.g., v10.x, v9.x). If the bridge registration pattern (E3), write-access permission flow (E4), or new-window routing (E5) differ materially from the inspected commit, then the paper's modeling requirements for Telegram are version-specific and cannot be assumed to hold cross-platform without a version-adaptation layer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's most concrete and novel claims about Telegram's security architecture (E3–E5 in Table 1) are derived from a manual reading of a single commit (62b56a07ca7e30e39f7fd00a6728d6bbd716ca1c) of the Telegram Android client. The authors explicitly acknowledge this scope limitation (Section 3, b96, b350). However, the load-bearing nature of this limitation is deeper than a typical scope note: the central argument that cross-platform analyzers must distinguish bridge implementations, operation-specific permission checks, and navigation contexts (Worked Example 2) rests entirely on the correctness and representativeness of this single-revision reading. If the inspected revision is atypical, or if the code paths have changed in subsequent releases (Telegram updates frequently), then the architectural distinctions that the paper elevates to modeling requirements may not hold for current or future Telegram MiniApps. The paper does not provide evidence that the inspected patterns are stable across versions, nor does it discuss the rate of change in the relevant code areas. This is not a flaw in the qualitative method per se, but it means the central claim about Telegram's architecture is empirically fragile: a single counterexample from a different revision would undermine the generality of the derived requirements.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper presents a qualitative, threat-informed systematization of the security architectures of WeChat Mini Programs and Telegram MiniApps, aimed at establishing a basis for cross-platform security analysis. WeChat is characterized from Tencent's official documentation and peer-reviewed studies; Telegram is characterized from official documentation plus a manual inspection of a single named revision of the Telegram for Android client (commit 62b56a07ca7e30e39f7fd00a6728d6bbd716ca1c). The authors define a shared taxonomy of security controls, assets, and threat sources, compare execution ecosystems and communication paths, and use two worked examples to derive two modeling requirements: distinguish the principal and validation boundary of each credential use, and associate host capabilities with concrete execution contexts and permission handlers. Table 1 traces the central claims to evidence locations, including file/line anchors in the inspected commit. The paper explicitly disclaims any security ranking of the ecosystems and any empirical detection evaluation.","tokens_in":17314,"tokens_out":11067,"duration_ms":81525,"significance":"If correct, the paper's central distinction—that platform mediation, credential-based authentication, web execution, and backend authorization protect different boundaries and should be modeled separately (Sec. 9)—gives designers of cross-platform MiniApp analyzers a concrete, actionable decomposition, and it correctly explains why existing WeChat-oriented tools (WEMINT, MiniCAT, TaintMini) do not transfer directly to Telegram. The paper's strengths are its evidence hygiene and scoping: Table 1 anchors claims to file/line locations in a named commit, the WeChat/Telegram evidence asymmetry is explicitly acknowledged (Sec. 3), source-backed statements are scoped to the inspected revision, and the authors repeatedly state that detection effectiveness is unevaluated (b332, b348, b358). The two worked examples are illustrative and structurally falsifiable. The main residual risk is the single-revision grounding of the Telegram-specific evidence (E3–E5); the authors acknowledge this limitation, and I assess it as substantially mitigated by the scoping and by the abstraction level of the derived requirements, as detailed in the major comment.","major_comments":[{"comment":"The derivation of modeling requirement 2, and Worked Example 2 in particular, is grounded in a manual reading of a single Telegram Android commit (62b56a07ca7e30e39f7fd00a6728d6bbd716ca1c). The authors properly acknowledge this (Sec. 3; b96, b350) and scope source-backed statements to the inspected revision (e.g., b259: 'In the inspected revision...'), which mitigates but does not resolve the concern. What remains unaddressed is the relationship between version-specific implementation details and the generality of the derived requirements: the paper never states which properties of the E3–E5 evidence (the dual-proxy registration under the same interface name, the write-access existing-authorization/confirmation branches, the ArticleViewer non-bot navigation path) are load-bearing for requirement 2 and which are illustrative, nor does it discuss the stability of these code areas across releases. I recommend adding a short paragraph in Sec. 8 that separates the structural invariant (an analyzer must resolve the concrete bridge implementation and navigation context before assigning host capabilities) from the version-specific instantiations, and that states explicitly what a counterexample from a different revision would and would not change. This would convert the acknowledged limitation into a precise falsifiability statement without expanding the paper's empirical scope.","section":"Sec. 7.4, Table 1 (E3–E5), Sec. 8 (Worked Example 2)","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The data availability statement says 'Table 1 provides source locations and permanent links to selected implementation paths,' but Table 1 as printed contains only file/line references and a commit hash, with no hyperlinks to those lines. Either add GitHub permalinks anchored to the cited lines or rephrase the sentence to say 'source locations' only.","section":"Data availability statement","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The control label 'S02 Secure Protocol' for MTProto asserts a security property that the paper does not establish and that is contested in the literature; consider renaming the control 'S02 MTProto' or 'S02 Encrypted Transport Protocol' to avoid endorsing the protocol's security.","section":"Sec. 7.2 / Table 3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Because the manual source inspection is the load-bearing evidence for the Telegram claims, the methodology would be more reproducible if it briefly described the analysis workflow (e.g., static navigation from BotWebViewContainer to nested classes, the specific search terms or call-graph steps used). A few sentences on tooling would strengthen the reproducibility of the E3–E5 findings.","section":"Sec. 3 (Methodology)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The layered permission model is described as 'adapted from the W3C MiniApp Standardization White Paper'; please state explicitly which elements come from the W3C model and which are the authors' additions, since Figure 5 is subsequently used as a comparative frame for the WeChat architecture.","section":"Sec. 6.2 / Figure 5","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Short version: this is a careful, honestly-scoped qualitative comparison of WeChat and Telegram MiniApp security architectures. It does not claim to be a detection study and says so repeatedly. The genuinely new material is the Telegram-side manual source reading—the bridge name TelegramWebviewProxy resolving to different concrete implementations depending on context, and the separate new-window routing path through ArticleViewer—which is concrete, tied to a named commit, and gives cross-platform tooling two specific modeling requirements: distinguish credential roles from validation boundaries, and associate host capabilities with concrete execution contexts and handlers.\n\nWhat it does well: the shared taxonomy (controls, assets, threats) is applied consistently on both sides; Table 1 traces each claim to an evidence location, including line numbers in a named commit; and the authors are disciplined about not ranking ecosystems by security. The distinction between platform mediation, credential-based authentication, web execution, and backend authorization is genuinely useful for anyone trying to build cross-platform analyzers.\n\nSoft spots, in proportion: the biggest is exactly what the stress-test notes—the Telegram findings rest on one manual reading of one Android commit. The authors acknowledge this (b350, b96), but acknowledgment does not change that E3–E5 are load-bearing for Worked Example 2. If the inspected revision is atypical or those code paths have moved, the derived modeling requirements lose their concrete grounding. A second revision or even a brief note on the stability of those code areas would have materially strengthened it. The WeChat side leans entirely on documentation and prior work—reasonable given closed source, but it means the comparison is only as good as its secondary sources. None of this is hidden; it is all stated, which is more than most papers this type manage.\n\nI largely agree with the reader: soundness 7 is about right, and the central architectural claims hold within their stated scope. The audience is people building MiniApp analysis tooling or entering the Telegram MiniApp security space and needing a map of trust boundaries. Send it to peer review; it is a solid, useful contribution in a young area, with traceability the community can build on.","headline":"A carefully scoped qualitative systematization whose Telegram-side bridge findings are concrete and new; the version-specific limits are acknowledged and the claims hold as scoped, so it deserves referee time.","tokens_in":17742,"tokens_out":1579,"duration_ms":10426},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"WeChat and Telegram hand different security jobs to the host and to developers; cross-platform MiniApp analyzers must model those boundaries separately.","keywords":["MiniApps","WeChat","Telegram","super-app ecosystems","security architecture","trust boundaries","Android WebView","JavaScript bridge"],"falsifier":"Inspect a later Telegram Android release or the iOS client at the same code locations: if TelegramWebviewProxy always dispatches through a single shared handler or if a bot new-window request passes through the same permission path as the SDK write-access handler, Worked Example 2's context-dependence claim fails for current versions. On WeChat, a documented flow where the session key directly authorizes server-side API calls, rather than only decrypting user data, would break the E1 credential-role split.","tokens_in":20397,"feed_emoji":"🔐","tokens_out":1719,"duration_ms":90768,"temperature":0.7,"pith_summary":"This paper tries to establish that MiniApp security is decided by which component actually enforces each control, the host platform or the developer, so cross-platform analyzers must model platform mediation, credential-based authentication, web execution, and backend authorization separately. WeChat runs packaged Mini Programs in a dedicated runtime with layered permission mediation and several distinct credentials, while Telegram loads plain web apps into a WebView and exposes host capabilities through a JavaScript bridge whose behavior changes with execution context. The paper's core warning is that a bridge name, token, or navigation event does not grant capability by itself: the same interface name maps to different handlers, and each credential protects a different boundary. Two worked examples turn this into concrete modeling requirements: credential-role analysis and context-dependent permission mediation. This matters because WeChat-only analyzers need architectural adapters before they can say anything useful about Telegram MiniApps, and neither platform's controls replace developer-side backend authorization.","feed_headline":"Same bridge name, different security in WeChat and Telegram MiniApps","feed_subtitle":"Platform mediation, credentials, and backend checks guard separate boundaries; cross-platform tools must model each one.","key_machinery":"The mechanism carrying the argument is a shared taxonomy of security controls, assets, and threat sources applied across execution environment, communication paths, and trust boundaries, with Table 1 mapping each selected claim (E1-E6) to its evidence and its implication for analysis. For Telegram, the load-bearing pieces are concrete code paths in the inspected Android client: bot-flag-dependent bridge registration in BotWebViewContainer, the operation-specific write-access handler with its existing-authorization and user-confirmation branches, and multi-window navigation through MyWebView, ArticleViewer, and onOpenUri. For WeChat, the machinery is the layered permission model adapted from the W3C MiniApp white paper plus the credential typology. The taxonomy makes the two ecosystems commensurable so the worked examples can turn architectural facts into requirements for future automated analysis.","core_discovery":"The paper's central claim is that the security responsibilities of a MiniApp ecosystem split between host platform and developer-operated components, and the split differs sharply between WeChat and Telegram. WeChat Mini Programs are packaged applications in a dedicated runtime: logic and render layers communicate through platform-managed channels, permissions are mediated in layers, and credentials (appID, AppSecret, access token, session key, login code) participate in different authentication flows. Telegram MiniApps are ordinary web applications loaded into an Android WebView, with host capabilities exposed through a local JavaScript bridge; in the inspected Telegram Android revision, the interface name TelegramWebviewProxy is registered in both bot and non-bot containers but dispatches to different handlers, write-access requests follow an operation-specific consent path, and new-window navigation takes a separate ArticleViewer and Browser route. From these architectures the paper derives two modeling requirements: distinguish the principal and validation boundary of each credential use, and associate host capabilities with concrete execution contexts and handlers. The paper scopes its Telegram findings to the inspected revision, does not rank the ecosystems by security, and does not claim validated vulnerability detection.","pith_inferences":["A cheap behavioral test follows from the E3 finding: dynamically compare event dispatch for the same bridge call in bot versus non-bot WebView contexts; if onEventReceived and onWebEventReceived do not differ, the context-dependence claim would need revision.","The credential-role schema (holder, principal, endpoint, operation) may generalize beyond WeChat and Telegram; applying it to Alipay or Baidu could test whether the taxonomy is a genuine cross-platform abstraction.","If platform tokens do not imply backend authorization, token-exposure scanners should pair exposure signals with reachability of protected operations, or they will over-report vulnerabilities in deployments whose backends still authorize.","Because the same bot token authenticates both Bot API calls and the HMAC integrity check over initData, an exposed token has two damage profiles; modeling separate flow edges would sharpen consequence analysis."],"forward_implications":["WeChat-only analysis tools need platform adapters that reconstruct execution contexts and bridge handlers before they can be applied to Telegram MiniApps.","Analyzers must trace each credential use to its validating principal and endpoint; a platform token, verified initData, or TLS encryption cannot by itself establish backend resource authorization.","A bridge name or window.open() event is not evidence of a permission bypass; the concrete bridge object, container mode, and operation-specific check path must be resolved first.","In both ecosystems, backend authentication and resource authorization remain developer responsibilities that no host-platform control replaces.","The taxonomy supplies a shared vocabulary and two modeling requirements that future automated cross-platform analyzers can be built on and evaluated against."],"supporting_citations":[{"why":"Inspected Telegram Android revision; BotWebViewContainer, ArticleViewer, and Browser code grounds E3-E5 on bridge, write-access, and navigation.","marker":"DrKLO/Telegram @ 62b56a07ca7e30e39f7fd00a6728d6bbd716ca1c"},{"why":"Supplies WeChat-side evidence E1 and E6 that application credentials and session material differ and IP allowlisting is configurable.","marker":"(Baskaran et al., 2023)"},{"why":"Official Mini Apps documentation grounds bot entry, SDK bridge, and initData validation procedure for E2 and A01.","marker":"(Telegram, 2026)"},{"why":"Documents the WeChat host environment with split logic and render layers, basis of the WeChat execution model.","marker":"(Tencent, 2026d)"},{"why":"Documents the access-token acquisition flow behind WeChat server-side authentication (S05, A03).","marker":"(Tencent, 2026c)"},{"why":"Closest prior WeChat-Telegram comparison; paper positions its component-level taxonomy against this attack-chain framing.","marker":"(Mohammadi et al., 2025)"},{"why":"MiniCAT packaged-app analyzer shows why WeChat-specific unpacking cannot transfer to Telegram's web model.","marker":"(Z. Zhang et al., 2024)"},{"why":"WEMINT WeChat static taint framework with large-scale data-leak findings characterizes the WeChat ecosystem side.","marker":"(Meng et al., 2023)"},{"why":"Provides the layered MiniApp permission model that the paper adapts for the WeChat comparison in Figure 5.","marker":"(W3C MiniApps Working Group, 2022)"}],"fun_headline_variants":["WeChat MiniApps run packaged while Telegram MiniApps run in a WebView","TelegramWebviewProxy reaches different handlers in bot and non-bot containers","Both ecosystems leave backend authentication and authorization to developers","Cross-platform MiniApp analysis needs credential roles and execution contexts","WeChat and Telegram split MiniApp security duties between host and developer"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The comparison assumes that one manual reading of Telegram Android commit 62b56a07ca7e30e39f7fd00a6728d6bbd716ca1c correctly captures bridge, write-access, and navigation behavior, and that Tencent's documentation plus prior peer-reviewed studies accurately describe WeChat's enforcement mechanisms.","fun_headline_variants_meta":{"raw":{"variants":["WeChat MiniApps run packaged while Telegram MiniApps run in a WebView","TelegramWebviewProxy reaches different handlers in bot and non-bot containers","Both ecosystems leave backend authentication and authorization to developers","Cross-platform MiniApp analysis needs credential roles and execution contexts","WeChat and Telegram split MiniApp security duties between host and developer"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"cost":0.000204113,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000204113,"upstream_inference_prompt_cost":8.6723e-05,"upstream_inference_completions_cost":0.00011739},"total_tokens":1598,"prompt_tokens":953,"completion_tokens":645,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":570}},"tokens_in":953,"tokens_out":645,"cost_source":"openrouter_usage","duration_ms":6703,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":570,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-11T11:06:55.104717+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-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":"Inspect a later Telegram Android release or the iOS client at the same code locations: if TelegramWebviewProxy always dispatches through a single shared handler or if a bot new-window request passes through the same permission path as the SDK write-access handler, Worked Example 2's context-dependence claim fails for current versions. On WeChat, a documented flow where the session key directly authorizes server-side API calls, rather than only decrypting user data, would break the E1 credential-role split.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies WeChat-side evidence E1 and E6 that application credentials and session material differ and IP allowlisting is configurable."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Closest prior WeChat-Telegram comparison; paper positions its component-level taxonomy against this attack-chain framing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MiniCAT packaged-app analyzer shows why WeChat-specific unpacking cannot transfer to Telegram's web model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"WEMINT WeChat static taint framework with large-scale data-leak findings characterizes the WeChat ecosystem side."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}