Back to skill

Security audit

中文视频知识抽取器

Security checks for vulnerabilities and agentic risk

Overview

This skill is mostly a coherent media-to-notes tool, but it can automatically upload private transcripts and local path metadata to any OpenAI-compatible endpoint configured in the environment without clear user consent or destination validation.

Install only if you are comfortable with the skill reading the media paths you provide, recursively processing folders, writing transcript and knowledge files under its output directory, and potentially sending transcript content and metadata to a remote LLM service whenever LLM_BASE_URL, LLM_API_KEY, and LLM_MODEL are set. Use a dedicated input folder, avoid sensitive recordings unless remote LLM variables are unset or controlled, and prefer trusted HTTPS endpoints with narrowly scoped API keys.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/process_media.py:868
Finding

Automatic Disclosure of Transcripts, Local Paths, and API Credentials to an Unrestricted LLM Endpoint

Content
View full analysis
bool: return bool(LLM_BASE_URL and LLM_API_KEY and LLM_MODEL) def normalize_llm_endpoint(base_url: str) -> str: base_url = base_url.rstrip("/") if base_url.endswith("/chat/completions"): return base_url return f"{base_url}/chat/completions" ``` ```python def call_openai_compatible_llm(messages: list[dict[str, str]]) -> str: url = normalize_llm_endpoint(LLM_BASE_URL) payload = { "model": LLM_MODEL, "messages": messages, "temperature": 0.2, } body = json.dumps(payload, ensure_ascii=False).encode("utf-8") request = Request( url, data=body, headers={ "Content-Type": "application/json", "Authorization": f"Bearer {LLM_API_KEY}", }, method="POST", ) try: with urlopen(request, timeout=LLM_TIMEOUT) as response: response_text = response.read().decode("utf-8", errors="ignore") except HTTPError as exc: raise RuntimeError(f"LLM request failed: {exc.code} {exc.reason}") from exc except URLError as exc: raise RuntimeError(f"LLM connection failed: {exc.reason}") from exc response_data = json.loads(response_text) choices = response_data.get("choices", []) if not choices: raise RuntimeError("The LLM response contains no choices") message = choices[0].get("message", {}) content = message.get("content" ...[truncated 3778 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
Findings (25)

Tainted flow: 'request' from os.getenv (line 882, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
97% confidence
Finding

The LLM endpoint is taken from environment configuration and transcript content is sent to it over the network, so whoever controls LLM_BASE_URL can redirect sensitive media-derived text and metadata to an arbitrary external service. In this skill's context, transcripts may contain confidential meeting content, private audio, or local file metadata, making exfiltration risk substantial.

Content

Scanner excerpt · scripts/process_media.py (reported line 893)May include surrounding context.

python
)

    try:
        with urlopen(request, timeout=LLM_TIMEOUT) as response:
            response_text = response.read().decode("utf-8", errors="ignore")
    except HTTPError as exc:
        raise RuntimeError(f"LLM 请求失败:{exc.code} {exc.reason}") from exc

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

文档写明在设置相关环境变量后,脚本会自动调用 OpenAI 兼容接口做后处理,但没有明确提醒用户这意味着输入内容或处理结果可能被发送到外部网络服务。对于涉及数据传输和隐私影响的行为,Markdown 描述应包含用户可见的警示。

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 46)May include surrounding context.

md
- 处理脚本:`scripts/process_media.py`

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The examples state that setting LLM-related environment variables will automatically enable LLM post-processing, but they do not disclose that transcripts, summaries, or extracted content may be sent to an external service. Because this skill processes potentially sensitive local media and transcripts, undisclosed outbound data transfer materially raises confidentiality and compliance risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill transmits transcript text and metadata to an external LLM endpoint without an explicit warning or consent mechanism. Given the tool accepts local files and folders, users may unknowingly send proprietary or private recordings off-host, making this a serious confidentiality issue.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

标题“中文视频知识抽取器”表明该技能面向中文,但 README 中未说明这是可选语言模式,也未提供其他语言选择或明确的区域/业务理由。根据规则,若技能在自然语言层面强制特定语言而无用户选择,可能构成语言/locale 政策问题。

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The README states that local folder input is supported and mentions recursive discovery, but it does not clearly warn users that providing a folder path may cause broad traversal of local media files and creation of multiple output artifacts. In an agent context, insufficient disclosure around recursive scanning and file generation can lead to unintended processing of sensitive local data and surprise writes to the filesystem.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
95% confidence
Finding

The skill advertises behavior that implies network access, local file reads/writes, shell/script execution, and possibly environment access, but it does not declare any explicit tool scope or permissions. That creates a transparency and least-privilege problem: an agent may invoke broader capabilities than the user reasonably expects, increasing the chance of unintended file access, downloads, or command execution.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The workflow says media may be downloaded from links and processed, but it does not warn that external content could be fetched or potentially transmitted to downstream processing components such as transcription or LLM enhancement. Users may unknowingly cause retrieval of remote content or disclosure of media-derived data to third-party services.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill states that local files and folders can be directly read and that fixed-format outputs will be generated, but it does not clearly warn users that their local content will be accessed and new files written. This can lead to inadvertent exposure of sensitive local data or unexpected modification of the workspace, especially when pointed at broad folders or mixed-content directories.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The display name is presented only in Chinese, which suggests a language-specific experience without any documented user opt-in or justification for a locale restriction. The file does not indicate that this skill is intended only for Chinese-speaking users or that alternate languages are supported.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The documentation explicitly says local folders will be recursively scanned for media files, but it provides no warning about scope, sensitive directories, or confirmation expectations. In a skill that accepts natural-language prompts and local paths, this increases the chance of unintentionally processing large or sensitive local content beyond what the user meant.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The trigger phrases are very broad, everyday Chinese requests such as asking to turn a video or folder into notes. In an agent environment, overly generic invocation language can cause accidental activation on routine conversations, which is more dangerous here because the skill can process URLs, local files, and recursive folders.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

This markdown template uses Chinese section headers and metadata labels throughout, which implies the generated notes will be structured in Chinese by default. Under the policy rule, forcing a specific language without user opt-in is a natural-language policy violation unless the locale restriction is clearly documented and justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The markdown template hardcodes Chinese headings and labels such as '来源', '一句话摘要', and '核心要点'. This imposes a specific language output format without any visible opt-in, alternative locale support, or justification that the skill is intentionally Chinese-only.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

Natural-language strings throughout the file, including the module description, prompts, logs, templates, and CLI help, are Chinese-only. The file does allow selecting transcription language, but it does not offer users any choice of interface or output language, which can violate language/locale policy if a fixed language is imposed without opt-in.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill reads LLM credentials from environment variables and uses them to send transcript data to a remote API, but this remote-processing capability is not evident from the declared behavior. That creates a security transparency issue: users may reasonably believe processing is local while their extracted content is actually transmitted externally.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/process_media.py (reported line 163)May include surrounding context.

python
def run_command(cmd: Sequence[str]) -> None:
    print("执行命令:", " ".join(cmd))
    subprocess.run(list(cmd), check=True)


def collect_media_files(folder: Path) -> List[Path]:

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill copies local media into its output tree and later writes transcripts, summaries, manifests, and JSON knowledge packs, creating additional persistent copies of potentially sensitive content. In a media-knowledge extraction tool, this materially increases exposure if the output directory is shared, backed up, or inspected by other users/processes.

Content

No source excerpt is available for this finding.

Ssd 1

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

Untrusted transcript text is embedded directly into LLM prompts, so spoken or embedded instructions in the media can manipulate the downstream model's behavior, output structure, or content prioritization. In this tool, that could taint generated notes, cause misleading summaries, or induce leakage of included metadata into outputs even if it does not directly execute code.

Content

No source excerpt is available for this finding.

Ssd 1

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The pipeline feeds chunk summaries derived from untrusted transcript content into a second LLM aggregation stage, allowing prompt-injection effects to persist and amplify across phases. This increases the chance that malicious spoken text shapes the final knowledge pack despite the intermediate summarization step.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The code sends transcript content and source metadata to an OpenAI-compatible API, but the skill description emphasizes extraction/summarization without explicitly disclosing outbound transmission. This omission increases the chance of unintentional data disclosure, especially for local files and private media sources.

Content

No source excerpt is available for this finding.

Ssd 1

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The full transcript plus timestamp preview are inserted directly into the final LLM prompt, maximizing the amount of attacker-controlled natural language that can steer the model. In a media ingestion workflow, adversarial audio could reliably bias summaries, produce deceptive chaptering, or inject policy-evading instructions into downstream knowledge artifacts.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
76% confidence
Finding

文件中的标题、说明和示例均默认要求以中文交互,但未向用户提供语言/locale 选择,也未说明该限制是否出于地区或合规原因。按照语言/locale 政策,这属于未获用户选择的单一语言约束。

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
95% confidence
Finding

This markdown template hardcodes Chinese field labels such as '来源', '类型', '时长', '语言', and '章节列表'. That imposes a specific language presentation without any visible opt-in, fallback, or documented region-specific justification, which matches the language/locale policy concern.

Content

No source excerpt is available for this finding.