Back to skill

Security audit

Huo15 Comic Edit

Security checks for vulnerabilities and agentic risk

Overview

The skill is advertised as a local FFmpeg video editor, but it ships under-disclosed cloud AI API helpers that can send prompts and local media to external services if invoked.

Review this package before installing if you expect a strictly local FFmpeg tool. The documented command appears local, but the package also includes unused cloud API helpers that can use ARK_API_KEY and transmit prompts, text, and selected local media to external services if another workflow calls them. Install only if you are comfortable with that extra bundled capability or can remove/disable the shared Ark helper.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (2)

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
scripts/_shared/ark_api.py:49
Finding

Undeclared Arbitrary Local File Upload Capability in a Local-Only Skill

Content
View full analysis
str: """Text-to-image / image-to-image. Returns the first image URL. reference_images accepts file paths or URLs and converts them to data URIs. """ content = [{"type": "text", "text": prompt}] for img in reference_images or []: content.append({ "type": "image_url", "image_url": {"url": self._image_to_data_uri(img)}, }) body = { "model": model or MODELS["image"], "content": content, "size": size, "watermark": False, } resp = requests.post( f"{self.base_url}/images/generations", headers=self.headers, json=body, timeout=120, ) ``` The local-file conversion is implemented as follows: ```python @staticmethod def _image_to_data_uri(path_or_url: str) -> str: if path_or_url.startswith(("http://", "https://", "data:")): return path_or_url p = pathlib.Path(path_or_url) ext = p.suffix.lower().lstrip(".") or "jpeg" if ext == "jpg": ext = "jpeg" b64 = base64.b64encode(p.read_bytes()).decode() return f"data:image/{ext};base64,{b64}" ``` ### Technical Analysis The Skill declares itself as a purely local FFmpeg editing utility. Its documented entry point, `scripts/edit.py`, does not require network access or remote image-generation capabilities. However, the bundled `ArkClient` accepts arbitrary caller-supplied paths through `reference_images`. `_image_to_data_uri()` resolves no trust boundary and directly reads the complete contents of each path using `p.read_bytes()`. The bytes are Base64-enco ...[truncated 2671 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Note
Location
scripts/_shared/ark_api.py:171
Finding

Unrestricted URL Download and Arbitrary Writable-Path Output

Content
View full analysis
pathlib.Path: out_path = pathlib.Path(out_path) out_path.parent.mkdir(parents=True, exist_ok=True) r = requests.get(url, stream=True, timeout=300) with open(out_path, "wb") as f: for chunk in r.iter_content(8192): f.write(chunk) return out_path ``` ### Technical Analysis `download_video()` accepts both the remote URL and local destination path without validation. The function performs a streaming GET request and writes the response directly to the selected path. The implementation does not: - Require HTTPS. - Restrict downloads to expected Volcengine media hosts. - Reject loopback, private, link-local, or metadata-service addresses. - Revalidate the destination after HTTP redirects. - Call `raise_for_status()` before writing the response. - Validate the response `Content-Type`. - Enforce a response-size or disk-usage limit. - Restrict `out_path` to the project directory. - Protect existing files from replacement. If untrusted values reach this method, the URL behavior can act as a server-side request forgery primitive from the execution environment. The output behavior can replace any file writable by the process. An unbounded response can also consume available storage. The reviewed documented entry point does not call this method, so exploitation requires another caller or future integration to pass attacker-controlled values. ### Attack Path #### Internal-resource access 1. An attacker controls or influences the `url` argument. 2. The attacker supplies a loopback, private-network, link-local, or redirecting URL. 3. `requests.get()` sends the request from the Skill's execution environment. 4. The response is wr ...[truncated 1548 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • 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
Findings (30)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The documented 0.3s fade effect appears unimplemented, and the actual input scope extends beyond the stated lipsync directory by falling back to videos/. While less severe than hidden network use, this is still a meaningful integrity issue because the skill may process different files than users expect and fails to deliver the declared transformation, reducing predictability and trust.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The documented 0.3s fade effect appears unimplemented, and the actual input scope extends beyond the stated lipsync directory by falling back to videos/. While less severe than hidden network use, this is still a meaningful integrity issue because the skill may process different files than users expect and fails to deliver the declared transformation, reducing predictability and trust.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The documented 0.3s fade effect appears unimplemented, and the actual input scope extends beyond the stated lipsync directory by falling back to videos/. While less severe than hidden network use, this is still a meaningful integrity issue because the skill may process different files than users expect and fails to deliver the declared transformation, reducing predictability and trust.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The documented 0.3s fade effect appears unimplemented, and the actual input scope extends beyond the stated lipsync directory by falling back to videos/. While less severe than hidden network use, this is still a meaningful integrity issue because the skill may process different files than users expect and fails to deliver the declared transformation, reducing predictability and trust.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The documented 0.3s fade effect appears unimplemented, and the actual input scope extends beyond the stated lipsync directory by falling back to videos/. While less severe than hidden network use, this is still a meaningful integrity issue because the skill may process different files than users expect and fails to deliver the declared transformation, reducing predictability and trust.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The documented 0.3s fade effect appears unimplemented, and the actual input scope extends beyond the stated lipsync directory by falling back to videos/. While less severe than hidden network use, this is still a meaningful integrity issue because the skill may process different files than users expect and fails to deliver the declared transformation, reducing predictability and trust.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The module exposes cloud image generation, video generation, downloading, and TTS capabilities that significantly exceed the manifest's declared FFmpeg-only local compositing purpose. This broad hidden capability increases attack surface, enables unexpected data egress and external service use, and undermines user/operator trust about what the skill actually does.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

Cloud image generation is unrelated to the stated FFmpeg concat/BGM/subtitle/fade task and can upload prompts plus local images to a third-party service. In a narrowly scoped editing skill, such undeclared functionality creates unnecessary exfiltration paths and materially expands what the skill can do beyond user expectations.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Remote AI video generation and polling are materially different from local FFmpeg compositing and allow submission of text, images, audio, and video to an external provider. Because the manifest frames this as a local editing helper, this hidden remote-generation path is more dangerous due to surprise data transfer, expanded capabilities, and potential misuse of provided credentials.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

Cloud TTS synthesis is outside the described FFmpeg-only editing workflow and sends user text to an external service using a credentialed API. In this skill context, undeclared voice synthesis increases data egress and capability scope without clear necessity, making it a genuine scope-deviation security concern.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
82% confidence
Finding

The skill documentation declares a purely local FFmpeg workflow, but the analyzer detected capabilities consistent with shell, file I/O, environment access, and possible network use without any explicit tool scope or permission boundaries. When a skill omits declared permissions, reviewers and users cannot reliably assess what the implementation may access, which increases the risk of over-privileged execution and hidden behavior.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The natural-language description specifies a Chinese visual style and Chinese font selection (国风样式, 宋体, Source Han Serif SC) but does not indicate that this is optional or limited to a China/Chinese-language use case. That can violate language/locale policy when the skill imposes a specific locale presentation without explicit user opt-in.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The natural-language documentation is entirely in Chinese and the TTS default voice is fixed to a Chinese voice profile, with no indication that language is configurable by user preference. This can be a language/locale policy issue when a skill implicitly forces a specific language without opt-in or documented regional scope.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The code loads an external API key from the environment and prepares authenticated access to a cloud AI endpoint despite the skill being presented as a local FFmpeg workflow. In this context, undisclosed network capability is dangerous because it permits external transmission and billing-sensitive operations that users would not reasonably expect.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
87% confidence
Finding

This performs authenticated external transmission of prompt and image data to a third-party image-generation API. External transmission is not inherently malicious, but in this skill it is under-declared relative to the manifest and therefore security-relevant due to privacy, compliance, and trust implications.

Content

Scanner excerpt · scripts/_shared/ark_api.py (reported line 66)May include surrounding context.

python
"size": size,
            "watermark": False,
        }
        resp = requests.post(
            f"{self.base_url}/images/generations",
            headers=self.headers,
            json=body,

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

These API calls transmit prompts and potentially attached media to an external provider, but this file provides no user-facing disclosure or consent mechanism. In a skill advertised for local FFmpeg compositing, the lack of transparency makes the external transmission materially more sensitive and likely to violate user expectations or policy.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

This submits externally facing video-generation requests that can include text and media inputs to a cloud service. Because the skill is described as local FFmpeg compositing, this undisclosed data egress is a real security concern even if the implementation intent is ordinary feature support.

Content

Scanner excerpt · scripts/_shared/ark_api.py (reported line 144)May include surrounding context.

python
if seed is not None:
            body["seed"] = seed

        resp = requests.post(
            f"{self.base_url}/contents/generations/tasks",
            headers=self.headers,
            json=body,

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The helper downloads arbitrary remote URLs to local files without host allowlisting, content validation, or even raise_for_status() checks. While this may be intended for normal media retrieval, in practice it can enable unexpected network access, storage of malicious or oversized content, and ingestion of untrusted files into downstream processing.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
87% confidence
Finding

This sends text to an external TTS service using authenticated API access. In isolation it is standard integration code, but within a skill presented as local-only compositing it becomes a genuine transparency and data-egress issue.

Content

Scanner excerpt · scripts/_shared/ark_api.py (reported line 207)May include surrounding context.

python
if emotion:
            body["emotion"] = emotion

        resp = requests.post(
            f"{self.base_url}/audio/speech",
            headers=self.headers,
            json=body,

Tainted flow: 'resp' from requests.get (line 158, network input) → pathlib.Path.write_bytes (file write)

Medium
Category
Data Flow
Confidence
65% confidence
Finding

Data from a source is assigned to a variable that is later passed to a sink, creating a variable-mediated taint flow.

Content

Scanner excerpt · scripts/_shared/ark_api.py (reported line 216)May include surrounding context.

python
resp.raise_for_status()
        out = pathlib.Path(out_path)
        out.parent.mkdir(parents=True, exist_ok=True)
        out.write_bytes(resp.content)
        return out

    # ------------ 工具 ------------

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The helper reads local files and base64-embeds them into data URIs for later upload, but there is no explicit notice here that file contents may leave the machine. In this skill context, silently packaging local media for external submission creates a meaningful privacy and data-handling risk.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The module docstring is written in Chinese, and the only natural-language documentation shown in the file uses that locale without offering any language choice or documenting a region-specific constraint. Under the stated policy, forcing a specific language without user opt-in can be a language/locale policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The sub_mark docstring contains Chinese-only instructional text and does not indicate that the user can choose another language. This appears to impose a specific language in natural-language documentation without visible opt-in or justification.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · scripts/_shared/config.py (reported line 35)May include surrounding context.

python
# API 端点(2026-04 核对)
ENDPOINTS = {
    "ark_base":      "https://ark.cn-beijing.volces.com/api/v3",
    "kling_base":    "https://api.klingai.com/v1",         # 官方;第三方 piapi.ai/fal.ai 亦可
    "suno_base":     "https://api.sunoapi.org",             # 第三方(Suno 无公开官方 API)
}

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

This code file contains natural-language docstrings and user-facing exception/warning messages exclusively in Chinese, beginning at the module docstring and continuing in runtime strings. The policy requires flagging language or locale constraints when the skill forces a specific language without user opt-in or documented justification.

Content

No source excerpt is available for this finding.