Back to skill

Security audit

本地视频字幕提取与翻译工具

Security checks for vulnerabilities and agentic risk

Overview

The skill does the advertised subtitle workflow, but it sends audio and transcripts to third-party AI services without clear user consent or tight scoping.

Install only if you are comfortable sending the selected audio and derived transcript to Groq and to the configured LLM provider, and run it in a limited environment with only the needed API keys. Avoid confidential, regulated, or highly private recordings unless provider terms and retention are acceptable.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (1)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:16
Finding

Unpinned Third-Party Dependencies Installed from a Mutable Package Source

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 16
Vulnerability Type: Unpinned runtime dependency installation
Risk Level: Medium

Vulnerable Code

bash
pip install requests openai

Technical Analysis

The installation instructions retrieve requests and openai without version constraints, integrity hashes, or a lock file. Consequently, the code installed depends on whichever package versions the package index resolves at installation time rather than versions reviewed together with this Skill.

This creates a supply-chain risk: a malicious or compromised future package release, a compromised package index or mirror, or unsafe package-index configuration could cause unreviewed code to execute during package installation or when the Skill imports these dependencies. The project does not itself retrieve and execute a remote script, but its mutable dependency resolution prevents the audited source from fully defining the code that will run.

Attack Path

  1. An attacker compromises a future release of one of the named dependencies, its publisher account, or a package source used by the environment.
  2. A user follows the documented pip install requests openai instruction.
  3. pip resolves and downloads the attacker-controlled or otherwise compromised package version because no reviewed version or hash is required.
  4. Malicious code executes during installation or when subtitle_skill.py imports and uses the package.
  5. The malicious dependency operates with the privileges of the Python or pip process and can access resources available to that process.

Impact Assessment

Successful exploitation could execute arbitrary Python code with the invoking user's privileges. In this Skill's expected environment, that scope may include access to local media and subtitle files, the GROQ_API_KEY and LLM_API_KEY environment variables, network connectivity, and other files readable or writable by the process. It could there ...[truncated 152 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace the mutable installation command with a reviewed, version-pinned dependency file, for example:

    text
    requests==REVIEWED_VERSION
    openai==REVIEWED_VERSION
    
  2. Generate and verify cryptographic hashes for all direct and transitive dependencies, then install with:

    bash
    python -m pip install --require-hashes -r requirements.txt
    
  3. Use a lock-file workflow so transitive dependency versions are reproducible.

  4. Explicitly use the trusted official package index or an organization-controlled, authenticated mirror.

  5. Run dependency vulnerability and provenance checks before updating locked versions.

  6. Install and run the Skill in an isolated virtual environment or container with minimal filesystem access and only the required environment variables.

Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Taint TrackingDirect Taint Flow, Variable-Mediated Taint Flow, Credential Exfiltration Chain
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (9)

Tainted flow: 'headers' from os.environ.get (line 44, credential/environment) → requests.post (network output)

Critical
Category
Data Flow
Confidence
90% confidence
Finding

Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.

Content

Scanner excerpt · subtitle_skill.py (reported line 50)May include surrounding context.

python
with open(audio_path, "rb") as f:
            files = {"file": f}
            data = {"model": "whisper-large-v3", "response_format": "verbose_json"}
            response = requests.post(url, headers=headers, files=files, data=data)
            response.raise_for_status()
    except Exception as e:
        return f"执行失败:调用 Groq 接口提取音频发生异常 - {str(e)}"

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The skill description materially understates behavior: it presents itself as local FFmpeg plus local Python processing, but the workflow sends audio and transcript content to external Groq/LLM services and does not actually produce the claimed bilingual-in-one-file output. This mismatch can defeat informed consent and security review, causing users to provide sensitive media under the false assumption that processing is local-only.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
94% confidence
Finding

The skill invokes external tools, reads environment variables, writes output files, and uses networked APIs, but it does not declare any explicit tool scope or permissions boundary. This creates an overbroad execution surface where an agent may use capabilities beyond what a reviewer or user expects, increasing the risk of unintended file access, secret exposure, or unreviewed outbound requests.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill instructs the agent to upload compressed audio to Groq Whisper and send derived transcript segments to an LLM for translation, but it does not prominently warn users that raw audio and transcript data leave the local environment. In a subtitle-processing context, this is especially dangerous because recordings may contain confidential conversations, PII, regulated content, or proprietary material, and users may reasonably expect local handling from the description.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The docstring states that the output will be translated into Chinese, and the translation prompt also instructs the model to translate subtitles into Chinese. This imposes a specific language/locale behavior without user opt-in or explanation that the skill is intentionally region-specific.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
89% confidence
Finding

The skill is configured to send subtitle content to an external LLM service using a base URL that is environment-controlled, with a default of DeepSeek. This means potentially sensitive transcript text leaves the local environment, and because the base URL is configurable, a misconfigured or malicious environment could redirect that data to an untrusted endpoint.

Content

Scanner excerpt · subtitle_skill.py (reported line 31)May include surrounding context.

python
# 1. 从环境变量动态获取配置 (提高安全性与环境适配性)
    GROQ_API_KEY = os.environ.get("GROQ_API_KEY")
    LLM_API_KEY = os.environ.get("LLM_API_KEY") # 你的 DeepSeek 等翻译大模型 Key
    LLM_BASE_URL = os.environ.get("LLM_BASE_URL", "https://api.deepseek.com/v1")
    LLM_MODEL = os.environ.get("LLM_MODEL", "deepseek-chat")

    # 校验环境与参数

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill uploads the local audio file to a third-party transcription API without any explicit user consent, warning, or data handling disclosure. Audio files can contain sensitive personal, business, or regulated information, so silent transmission to an external processor creates a real confidentiality and compliance risk.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
94% confidence
Finding

The code sends the full local audio file to Groq's external transcription API. In the context of a subtitle extraction tool, this behavior is functional, but it still creates a real privacy and compliance risk because raw audio often contains sensitive information and there is no visible consent or minimization step.

Content

Scanner excerpt · subtitle_skill.py (reported line 43)May include surrounding context.

python
print(f"[Skill 运行中] 开始处理音频: {audio_path}")

    # 2. 调用 Groq (Whisper) 提取时间轴
    url = "https://api.groq.com/openai/v1/audio/transcriptions"
    headers = {"Authorization": f"Bearer {GROQ_API_KEY}"}
    
    try:

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
83% confidence
Finding

The function accesses GROQ_API_KEY and LLM_API_KEY from the environment, which is sensitive configuration data. While this is common practice, the file provides no explicit warning or explanatory notice to users that the skill depends on and accesses credentials from environment variables.

Content

No source excerpt is available for this finding.