AI Subtitle Translator

AI Security Review

Evidence context: threat category none; evidence quality moderate.

AI Subtitle Translator (developer Vinner Vi, version 2.7.1, Firefox, 24 users) advertises live subtitle translation for YouTube, Bilibili, TikTok and roughly fifty other video sites in thirty-plus languages, on a credit model with 50 free translations. Nothing in the scan contradicts that description. Across 481 findings there are zero malware signatures, zero obfuscation matches, zero exposed secrets, zero dependency hits and zero tool-poisoning hits. What remains is network and code-quality noise around code any working translator would contain.

The 38 network findings are fetch call sites: NET-FETCH-auth.js-285, NET-FETCH-auth.js-381, NET-FETCH-auth.js-496, NET-FETCH-auth.js-532, NET-FETCH-popup/popup.js-5284, five entries in background/service-worker.js between lines 598 and 1566 plus three more near line 2900, NET-FETCH-content/content.js-1587 and NET-FETCH-content/primevideo-interceptor.js-645. A translation product has to call a backend to translate, and it has to call an identity provider to meter credits. The 744453102252-6p9fgjl10hs7h5njuo81m9mav31i3p7a.apps.googleusercontent.com client IDs sit next to those calls, which fits a Google sign-in flow for the credit balance. content/primevideo-interceptor.js matches the feature set: Prime Video serves subtitles from its own player, so lifting the text means hooking the player's media requests.

The 291 indicator entries look large in bulk and are not. The endpoints include aljazeera.com, alison.com, amazon.co.uk, amazon.co.jp, academy.scienceforsport.com, academy.inkmentors.com and streaming hosts such as 123av.com and alloha.yani.tv. That reads as the supported-site allowlist the listing itself promises, and the content script has to recognise each host before it runs. Several of the streaming hosts (ag2m4.cfd, 8tsu.net) are piracy-adjacent, which is a content-policy question rather than a security one. No finding anywhere in the scan pairs a cookie or credential read with a post to any of these hosts.

The 151 code-smell matches are all low severity, and not one shares a file with a malware signature. That is the shape this scanner produces on any bundled service worker, not a detection. A skeptic would point at auth.js holding Google OAuth client IDs while background/service-worker.js makes repeated network calls, and would note the single manifest-analysis finding plus empty permission and host-permission arrays in the data we could read. That gap is real: we cannot confirm from this bundle which origins the manifest grants at install time. It still does not tip the verdict, because a malicious package would need somewhere to send data, and the only destinations here are a site allowlist, Google identity endpoints, and the vendor's own translation service.

There is no malware signature, no obfuscated payload, no credential access and no hijacked search or new-tab behaviour in the evidence. The finding volume is an artefact of domain extraction and broad JavaScript rules.

Key Reasons

• Zero malware signatures, zero obfuscation, zero secrets and zero tool-poisoning matches across 481 findings
• The 291 indicators are a supported-site allowlist (aljazeera.com, amazon.co.uk, alison.com, streaming hosts) that matches the advertised 50+ site coverage
• NET-FETCH hits such as auth.js:285 and background/service-worker.js:598 are ordinary API and sign-in calls required by a translation service with Google OAuth client IDs
• 151 code-smell matches are all low severity and none shares a file with a malware signature
• Named developer, versioned release 2.7.1, no browser-hijack or credential-access behaviour in the evidence

False Positive Considerations

• Domain extractor counting the extension's supported-site list as 291 indicators
• Low-severity code-smell rules firing on a bundled background service worker
• NET-FETCH rule firing on every ordinary fetch call site in auth.js, popup.js and service-worker.js
• Medium-severity labelling applied to network call sites with no destination analysis

Reviewed 2026-10-18; recommended action: suppress false positive; model confidence 78%.