Security audit
Security checks for vulnerabilities and agentic risk
This movie recommendation skill is coherent and purpose-aligned, with disclosed Douban lookups and local movie-list storage but some privacy and reliability caveats.
Before installing, understand that movie searches may be sent to Douban and that cache, watched-history, and wishlist JSON files may be stored locally. Avoid recording sensitive viewing preferences on shared machines, and review or delete the generated JSON files if you do not want that history retained. The watched/wishlist command parsing is also incomplete and may save placeholder entries.
The documentation describes real-time Douban lookups and crawling/API usage without warning that user queries may be transmitted to an external service. Even simple movie searches can reveal user interests, language, or behavior patterns, and undisclosed third-party transmission is a meaningful privacy issue.
The README advertises viewing-record and wishlist management but does not clearly disclose that movie history and user ratings are persisted locally. This can mislead users into sharing behavioral data without informed consent, creating privacy and retention risks if the files are later exposed or accessed by other components.
The example trigger phrases are ordinary conversational requests like asking for movie recommendations, which creates a real risk of over-broad or unintended activation in a host assistant. In a multi-skill environment, generic phrasing can cause this skill to intercept unrelated user requests and process them without clear user intent to invoke it.
The trigger phrases are very broad and map to ordinary movie-related conversation, so the skill may activate in situations the user did not explicitly intend. In a skill with network access and local file write capability, ambiguous invocation increases the chance of unintended scraping, recommendation actions, or persistence of viewing data without clear user consent.
The skill explicitly requests network access, web scraping, and local file storage, but it does not disclose what user data may be sent externally, scraped, stored locally, or retained over time. This creates privacy and transparency risks, especially because movie preferences, ratings, and comments can reveal sensitive interests or habits and may be persisted without informed consent.
The file's natural-language descriptions and all user-facing responses are written in Chinese, and the trigger/interaction model assumes Chinese input terms such as moods and commands. There is no opt-in, language selection, or documented justification that this skill is intended only for a Chinese-language locale.
The skill persists watch history and watchlist data to local JSON files without any user-facing disclosure, consent prompt, or retention notice. In an agent setting, silent storage of user preferences and activity can expose sensitive behavioral data, especially on shared systems or where users do not expect persistent logging.
Movie lookups send user-provided queries to an external Douban-backed API/search layer without explicitly informing the user that their input is transmitted off-device. User queries can reveal preferences, interests, or personal context, so undisclosed external transmission is a privacy risk in this skill context.
The help text says users can record a specific movie and rating (for example, "我看了 XXX，评分 5 分"), but the implementation does not parse the movie name or rating at all. Instead it always calls add_watched("未知电影", comment="自动记录"), contradicting the stated behavior rather than merely omitting a minor detail.
The default help advertises functionality like "把 XXX 加入想看", implying the skill will add the requested title. However, the implementation ignores the query content and always saves "未知电影", which directly conflicts with the described capability.
All user-facing instructions, examples, and outputs are presented in Chinese, which effectively imposes a single language without documenting an opt-in or locale-specific scope. Under the language/locale policy, this is a natural-language constraint unless the skill clearly offers alternatives or justifies a China-specific audience.
The request headers hard-code Accept-Language to prefer zh-CN, which imposes a specific language/locale behavior in the skill's network interactions. This matches the policy category for language or locale constraints because the file does not indicate user opt-in or a documented region-specific necessity.
The module-level documentation states the implementation uses a web-scraping approach and does not rely on an API, while this docstring says the fallback is used 'when the Douban API is unavailable'. That is an active contradiction in the code documentation about what upstream mechanism the skill actually depends on.
The function get_hot_movies is documented as fetching popular/currently showing movies, but it directly calls get_top250(start=0), which is a ranked Top250 list rather than a hot/in-theaters feed. This is an active contradiction between the inline documentation/name and the implemented behavior.
No suspicious patterns detected.
