Back to skill

Security audit

带货广告视频自动出片(配音混剪全包)

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because scanners report undisclosed cloud processing, duplicate-detection evasion, and other mismatches or safety issues alongside its video-advertising features.

Treat the Review verdict as a reason to pause before installing. Consider whether you accept external processing of narration, copy, and campaign metadata; local editing-project discovery; and the duplicate-detection evasion capability. The scan inspected 120 of 122 components, with two partially inspected, and the tool-poisoning analysis was degraded, so the evidence is incomplete.

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

Warning
Location
engine/adflow/agents/match_agent.py
Finding

Cloud matching fallback bypasses the documented no-external-LLM workflow

Content
View full analysis

Impact

Material tags, scene descriptions, identifiers and performance metadata can be disclosed to the configured cloud provider, including metadata for candidates not ultimately selected. Requests can also consume paid API quota. This is not evidence that API keys are sent to an unrelated recipient, nor that raw video files are uploaded by this fallback.

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Taint TrackingDirect Taint Flow, Variable-Mediated Taint Flow, Credential Exfiltration Chain
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (139)

Tainted flow: 'headers' from os.environ.get (line 127, 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 · engine/adflow/engine/minimax_tts.py (reported line 135)May include surrounding context.

python
logger.info("MiniMax TTS: %s voice=%s", text[:40], self.voice)

        try:
            r = requests.post(self.API_URL, json=body, headers=headers, timeout=30)
            data = r.json()
        except Exception as e:
            raise RuntimeError(f"MiniMax API 请求失败: {e}")

Tainted flow: 'request_body' from os.environ.get (line 144, 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 · engine/adflow/engine/volcano_tts.py (reported line 177)May include surrounding context.

python
text[:40], self.voice, emotion, speed_ratio)

        try:
            response = requests.post(
                self.API_URL,
                json=request_body,
                headers={

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Narration synthesis and local audio processing directly support the declared video-generation purpose and are not themselves mismatches. However, this backend raises a RuntimeError unless VOLCANO_TTS_APP_ID and VOLCANO_TTS_TOKEN (or equivalent configuration values) are provided, then authenticates requests to https://openspeech.bytedance.com/api/v1/tts. That materially contradicts the explicit no-API-key claim. The supplied chunk does not establish whether a credential-free alternative exists elsewhere; missing video-editing or advertising-law-checking code in this isolated TTS module is not treated as an additional mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Narration synthesis and local audio processing support the declared video-generation purpose and are not themselves mismatches. However, this backend explicitly raises an error when no API key is configured and uses that key for requests to https://api.minimax.io/v1/t2a_v2. This materially contradicts the no-API-key claim. The supplied chunk does not establish whether other keyless backends exist or whether this backend is optional; the finding concerns the implementation shown. Features implemented elsewhere, such as video editing or prohibited-word checks, are not treated as missing based on this partial code.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

Ordinary editing and variant generation support the declared batch-video workflow. However, this module specifically targets making near-duplicate variants pass platform duplicate-detection systems, a purpose not disclosed in the description. The executable behavior supports that distinction: it applies small visual alterations and compares perceptual and textual similarity. The chunk does not implement all seven dimensions claimed in its comments, nor does it establish that detection evasion succeeds. Local video processing and FFmpeg execution are consistent supporting resources; the mismatch is the additional anti-detection capability, not the absence of the rest of the advertised workflow from this single module.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

Most functionality is directly consistent with the advertised subtitle and video-production workflow. However, _llm_split_to_n retrieves an API key, base URL, and model, then submits the supplied text through OpenAI chat completions when a key exists. This introduces an undisclosed outbound text-processing path and API credential use. The API is optional and local fallbacks exist, so the code does not establish that keys are required; the mismatch concerns the unqualified API-key-free presentation and undisclosed remote processing. Other advertised features being absent from this subtitle-specific chunk is not itself a mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

Style import could support the broader video-generation workflow, but automatically searching existing editor projects and reading the newest draft is a distinct, undisclosed local-data access capability. Despite its interactive naming, the discovery function does not ask the user to choose a draft. The supplied chunk performs style extraction, not the advertised generation or advertising-compliance checks; their absence here does not establish that the broader skill lacks them. No network transmission, file modification, or API-key use is shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

Text segmentation is a legitimate supporting component of subtitle generation; this module need not implement the entire advertised video workflow. The material discrepancy is its cloud fallback: available() obtains an API key, endpoint, and model from get_llm_config(), and _model_segment() sends the supplied narration to that endpoint. This introduces an undisclosed external processing path and qualifies the no-API-key claim. Cloud access is conditional, not mandatory: local inference and punctuation fallback can operate without a cloud API key.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

LLM calls, retries, usage tracking, and JSON parsing are reasonable supporting details for the advertised video workflow and are not independently mismatches. However, the explicit claim that no API key is needed is not supported by this implementation: its primary client uses API credentials, while local operation is optional and treated as a fallback. Credentials could potentially be supplied by the service rather than the user, but that provisioning is not established by this chunk. The absence of video rendering or advertising-law checks in this base-agent module alone does not demonstrate that those features are missing from the full skill.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

The supplied module implements a substantial advertising-analytics and training-data workflow not disclosed in the description, rather than merely storing intermediate video-generation artifacts. Its CSV ingestion, performance-label management, and training-data preparation extend beyond the stated video-production purpose. The absence of generation or prohibited-term checking in this individual module does not establish that those features are absent elsewhere. No network transmission or actual model training is shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

Most behavior supports the declared batch-video production purpose. However, _shuffle_shot_order explicitly describes reordering shots to avoid duplicate detection, and _build_post_vf invokes AntiDedupEngine with enabled defaulting to true. This introduces a detection-evasion objective beyond ordinary editing or generating creative variants. The supplied chunk also does not demonstrate copywriting, narration generation, or prohibited-word checking: it consumes existing scripts and narration and marks outputs pending_review with compliance_check_passed=False. Those omissions alone do not establish a mismatch because this is only one component of the skill.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
87% confidence
Finding

LLM configuration is a reasonable supporting implementation detail for generating advertising copy. The mismatch is specifically the no-API-key claim: the default provider is DashScope, with a DASHSCOPE_API_KEY lookup, and several other providers likewise use configured or environment-supplied credentials. Local Ollama offers a potentially keyless alternative, but this chunk does not establish a universally keyless workflow. There is no evidence here of credential exfiltration or unrelated behavior, and the absence of video-generation functions in this supporting module is not itself a mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
86% confidence
Finding

The supplied module implements a distinct analytics/reporting capability that the description does not disclose. Although its video-related factors could support a broader advertising-video workflow, its actual function is analyzing previously labeled videos rather than producing advertising videos. No suspicious network access, exfiltration, or unrelated resource use is evident. This assessment is limited to the supplied module and does not establish whether other components implement the advertised creation features.

Content

No source excerpt is available for this finding.

Mandatory plain-language rewriting introduces unsupported health and safety assurances.

High
Category
Prompt Injection
Confidence
99% confidence
Finding

Plain-language rewriting is explicitly required to introduce stronger health and safety claims than the supplied technical facts support. This is the same underlying defect as the overlapping SDI-1 finding, with potential consumer harm and regulatory exposure despite the prompt's separate prohibition on false promises.

Content

No source excerpt is available for this finding.

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · engine/adflow/agents/script_agent.py (reported line 480)May include surrounding context.

python
except Exception:
            pass

        return prompt

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · engine/adflow/config_loader.py (reported line 54)May include surrounding context.

python
with open(config_path, "r", encoding="utf-8") as f:
            self._data = _yaml_rt.load(f) or {}

        # 加载 .env 文件
        self._load_dotenv()

        # 环境变量覆盖(最高优先级)

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · engine/adflow/config_loader.py (reported line 63)May include surrounding context.

python
@staticmethod
    def _load_dotenv() -> None:
        """加载项目 .env 文件到 os.environ"""
        env_path = Path(__file__).parent.parent / ".env"
        if not env_path.exists():
            return

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · engine/adflow/config_loader.py (reported line 64)May include surrounding context.

python
@staticmethod
    def _load_dotenv() -> None:
        """加载项目 .env 文件到 os.environ"""
        env_path = Path(__file__).parent.parent / ".env"
        if not env_path.exists():
            return
        try:

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · engine/adflow/config_loader.py (reported line 79)May include surrounding context.

python
if key and value and key not in os.environ:
                        os.environ[key] = value
        except Exception:
            pass  # .env 加载失败不阻塞启动

    @staticmethod
    def save_env(updates: dict[str, str]) -> None:

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/engine-reference.md (reported line 96)May include surrounding context.

md
adflow/            引擎源码(agents/engine/db/gui/compliance/...)
  main.py            可选图形界面(python main.py)
  config.yaml        全局配置
  .env               付费配音 key(用户自建,不随包分发)
  materials/         视频素材库
  bgm_library/       背景 Music(用户自有授权)
  data/              数据库+关键帧

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · scripts/check_env.py (reported line 121)May include surrounding context.

python
if backend == "edge":
    ok("配音方案: Edge-TTS(微软免费,无需任何 key)")
elif backend in ("volcano", "minimax"):
    # 检查 .env 里有没有配 key
    env_text = ""
    env_path = os.path.join(ENGINE, ".env")
    if os.path.isfile(env_path):

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · engine/adflow/config_loader.py (reported line 83)May include surrounding context.

python
@staticmethod
    def save_env(updates: dict[str, str]) -> None:
        """把 key=value 写回 .env 并同步 os.environ(设置页保存 API Key 用)

        只更新给了非空值的键;没提到的键原样保留。
        """

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · engine/adflow/config_loader.py (reported line 87)May include surrounding context.

python
只更新给了非空值的键;没提到的键原样保留。
        """
        env_path = Path(__file__).parent.parent / ".env"
        lines: list[str] = []
        if env_path.exists():
            try:

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · engine/adflow/gui/settings_panel.py (reported line 120)May include surrounding context.

python
"在顶部搜索框搜「语音技术」,进语音技术控制台",
                "开通「语音合成大模型」服务",
                "点「应用管理」→ 创建应用 → 复制 App ID",
                "在应用详情里点「获取访问令牌」→ 复制 Access Token",
                "两个一起填到上面两个框 → 保存",
            ],
        },

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/tts-setup.md (reported line 26)May include surrounding context.

  1. 打开 https://www.volcengine.com/ 注册并完成实名认证(个人可)
  2. 控制台搜「语音技术」→ 开通 语音合成大模型声音复刻/语音合成(按字符计费,新用户一般有免费额度)
  3. 在控制台拿到两个值:App ID(应用 ID)和 Access Token(访问令牌)
    • 拿令牌的位置一般在「语音技术 → API Key 管理」;建应用后看应用详情页
  4. 用记事本打开(没有就新建)engine/.env,加两行(换成你自己的值):
    text