T08 · Insecure Dependencies
- Location
SKILL.md:23- Finding
Unpinned Third-Party Package Installation
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a local Whisper transcription helper with some overstated feature and offline claims, but I found no hidden exfiltration, destructive behavior, persistence, or deceptive execution.
Install only if you are comfortable with a Chinese-focused transcription workflow that may rewrite or summarize text unless you ask for raw transcription. For stricter privacy or reproducibility, preinstall and pin `openai-whisper` in a virtual environment, download the Whisper model intentionally, and treat 'fully offline' as applying only after dependencies and model weights are already present.
SKILL.md:23Unpinned Third-Party Package Installation
声明描述的核心能力是离线转写与文本整理总结,而实际代码仅是一个模型下载工具。它接收模型名称和可选下载目录,调用 whisper.load_model 将模型下载/缓存到本地。代码的主要目的与声明的主要用途存在实质性偏差:声明的是完整的本地语音转文字与后处理功能,实际却只是前置的模型准备步骤。此外,声明强调“全程离线”,但该脚本下载模型通常需要网络访问,这一点也与描述不一致。尽管模型下载可被视为支持性实现细节,但当前提供的代码块本身并未体现声明中的核心功能,因此应判定为描述与行为不匹配。
声明描述的核心功能不仅包括本地离线 Whisper 转写,还包括对结果进行进一步加工,生成清理文本、书面整理稿和结构化总结稿。但实际代码只做了 Whisper 模型加载、设备选择、音视频文件转写/翻译,以及将 result['text'] 原样保存为单个 txt 文件,没有任何文本清洗、改写、结构化总结、多版本输出或后处理逻辑。因此描述明显高估了代码能力,存在实质性功能不一致。另一个较小差异是代码支持 task=translate 的翻译模式,而声明仅提到转文字;不过主要问题仍是声明中的多种生成能力未实现。
声明描述的是一个完整的本地离线语音转写与内容整理技能;但提供的代码块只是一个模型下载工具。它通过 whisper.load_model 下载并缓存模型,主要用途是准备依赖资源,而不是执行转写流程。尤其声明中的“全程离线”与该脚本的下载行为明显不符,因为下载模型通常需要网络访问。虽然下载模型可被视为相关配套步骤,但此代码块本身与声明的核心功能存在实质性差异,因此应判定为描述与行为不匹配。
代码的核心行为确实包含“本地 Whisper 语音转文字”以及“支持音频和视频输入”,这一部分与描述基本一致。但描述中承诺会“自动生成清理文本、书面整理稿和结构化总结稿”,而代码实际只调用 Whisper 的 transcribe/translate 得到单一文本结果,并写入一个 txt 文件,没有任何文本清洗、整理、总结或结构化输出逻辑。此外,描述称“全程离线”,但代码通过 whisper.load_model(..., download_root=...) 加载模型;若本地不存在模型,Whisper 可能联网下载,因此该表述也不完全准确。综上,声明的功能范围明显大于实际实现,存在实质性描述与行为不符。
The skill description says it will generate '清理文本、书面整理稿和结构化总结稿' in addition to transcription. The implementation only extracts result['text'] from Whisper and writes a single plain text file, with no cleaning, editorial rewrite, or summarization logic present.
技能清单声明“全程离线”,但此脚本通过 whisper.load_model 在本地无缓存时获取模型,实际依赖网络下载。该行为超出了“本地 Whisper 语音转文字、全程离线”的语义范围,因为代码并非仅处理本地音视频和本地模型。
The README specifies Chinese-only behavior such as '简体' output and rewriting spoken language into polished written Chinese, but it does not offer the user any language or locale choice. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation.
The skill describes behaviors that write files to user-controlled paths and generate multiple output artifacts, but it declares no explicit tool scope or permissions boundaries. In an agent environment, missing scope makes it harder to constrain where writes occur or what file operations are permitted, increasing the chance of unintended file modification.
Forcing Traditional-to-Simplified conversion without user opt-in can alter meaning, names, quotations, or compliance-sensitive records, especially for transcripts that are expected to preserve source fidelity. In a transcription skill, silent normalization is risky because users may rely on outputs as accurate records of spoken content.
Mandating third-person rewriting, logical reorganization, compression, and stylistic transformation without user choice can materially change the substance and attribution of the original speech. In this skill context, that is dangerous because users may mistake rewritten content for faithful transcription, creating integrity, legal, or reputational risks.
The user-facing natural-language strings in the command description, help text, and errors are written only in Chinese. This imposes a specific language on users without any opt-in or alternative, which matches the language/locale policy violation criteria.
The skill description promises fully offline processing, but whisper.load_model(..., download_root=args.download_dir) will download model files if they are not already present. In a privacy-sensitive transcription tool, this can violate user expectations, trigger unintended network access, and leak metadata about model usage or environment state.
The skill description says it automatically generates '清理文本、书面整理稿和结构化总结稿', implying multiple postprocessed outputs beyond transcription. This script only runs Whisper transcription/translation and writes result['text'] to a single txt file, with no cleanup, formatting, or summarization logic.
The manifest says the skill operates '全程离线' and describes local Whisper processing, but the code exposes a --download-dir option and passes it to whisper.load_model(), which will download model weights if they are not already present locally. That behavior is broader than a strictly offline-only implementation.
The manifest includes the keyword "chinese", which can imply a language-specific behavior or locale targeting, but the file does not explain whether Chinese is optional, user-selected, or required for a justified regional use case. Because policy checks apply to natural-language metadata too, this could be interpreted as a language/locale preference without explicit opt-in.
The script's error, help, and status messages are all hard-coded in Chinese, including the dependency error text and argparse descriptions. This imposes a specific language on all users without any opt-in or alternative locale handling, which matches the language/locale policy violation category.
The script emits user-facing text in Chinese, including the installation error message, which imposes a specific language on users without opt-in. This matches the locale policy concern because the file provides no alternative language selection or justification for being Chinese-only.
The argparse description, help text, and progress/status prints are all Chinese-only, and the script does not offer any language or locale option. For a general-purpose utility, this is a natural-language policy issue unless the locale restriction is explicitly documented and justified.
The argument parser description, help strings, and error message are written only in Chinese, which imposes a specific language on all users. The file does not offer a locale choice or indicate that the tool is intentionally restricted to a Chinese-speaking audience.
The script calls whisper.load_model with a configurable download directory, and if the requested model is not already present, Whisper may download it. There is no confirmation prompt or explicit user-facing warning before this network activity occurs, so users may not realize the tool can contact external services.