Back to skill

Security audit

Youtube数据助手

Security checks for vulnerabilities and agentic risk

Overview

ClawScan’s verdict is benign: the skill is presented as a read-only YouTube analysis tool, and the reported concerns are about consent and disclosure, persistent local records, broad triggers, and an unpinned installer rather than clear intentional harm.

The verdict is benign, but that does not mean risk-free. Before installing, consider whether you are comfortable with the reported third-party sharing without a privacy warning, persistent local compliance and audit records, Chinese-only consent wording, and an unpinned installer. Skillspector marked its analysis partial because some references could not be resolved, so coverage is incomplete.

Vulnerability Patterns
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (16)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Ordinary supporting logs would not alone establish a mismatch. Here, however, the code explicitly marks legal version 1.0.0 as confirmed and records a first_time_confirmation event without any confirmation interaction or prerequisite check. This changes compliance state rather than merely logging YouTube analysis. The mismatch is that unsupported confirmation behavior, not simply the absence of API calls in this limited chunk.

Content

No source excerpt is available for this finding.

The consent gate requires Chinese-language confirmation and explicitly rejects English confirmation without offering a language choice.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

All accepted confirmation phrases are Chinese, while English “yes” is explicitly rejected as ambiguous. This imposes a language constraint on users without opt-in or a documented justification for requiring Chinese consent; citing Chinese laws does not itself explain that restriction.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skillhub'.

Medium
Category
MCP Rug Pull
Confidence
88% confidence
Finding

The installation command invokes an unpinned npm executable, allowing the code executed during installation to change without review. A compromised future package release could execute with the installing user's privileges; this is a conditional supply-chain exposure, although the referenced executable is an installer rather than an MCP server.

Content

No source excerpt is available for this finding.

The consent gate requires Chinese-language confirmation and explicitly rejects English assent without offering a language choice.

Medium
Category
Output Handling
Confidence
98% confidence
Finding

L321 accepts only listed Chinese confirmation phrases, while L322 explicitly rejects “yes” and L323 directs users back to Chinese wording. This imposes a language requirement on a general-purpose YouTube skill without user opt-in; requiring explicit legal consent does not justify restricting consent to Chinese.

Content

No source excerpt is available for this finding.

Channel lookup workflows send user-supplied channel details to a third-party API without a privacy warning.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

These chain recipes instruct API calls using user-provided channel URLs or identifiers, rather than merely listing available tools. The API host is www.aconfig.cn (L003), not YouTube, but the file does not warn users that their lookup details will be disclosed to that third-party service.

Content

No source excerpt is available for this finding.

L300–L306 authorize a MaxHub account-balance query outside the stated public YouTube analytics scope.

Medium
Category
Excessive Agency
Confidence
92% confidence
Finding

The whitelist includes GET /api/maxhub/balance, exposing an account-level balance capability rather than public YouTube content or channel data. The manifest does not declare account or billing management, and querying that balance is not a direct requirement of the described analytics operations.

Content

No source excerpt is available for this finding.

The mandatory unsupported-request response forces Chinese output without user language opt-in.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

L051 instructs the agent to reply directly with a fixed Chinese sentence whenever a request is unsupported. This mandates Chinese user-facing output without offering a language choice or justifying a locale restriction.

Content

No source excerpt is available for this finding.

Broad standalone triggers can activate video recipes for unrelated requests.

Medium
Category
Trigger Abuse
Confidence
97% confidence
Finding

Triggers such as “趋势” (trends), “看看评论” (look at comments), and “看看回复” (look at replies) also occur in requests unrelated to YouTube or video content. The matching rules resolve competing recipe matches but do not require a relevant platform context or target, so these phrases can cause unintended recipe invocation.

Content

No source excerpt is available for this finding.

The video-comment recipe uses generic reply-related activation phrases without requiring video context.

Medium
Category
Trigger Abuse
Confidence
94% confidence
Finding

The phrases “评论回复” (“comment replies”) and “看看回复” (“look at the replies”) can refer to posts, messages, or other conversations, not specifically video comments. Although the recipe requires a video_id, its activation condition does not explicitly require an identified video or exclude non-video contexts.

Content

No source excerpt is available for this finding.

The search workflow sends user search terms to a third-party API without a privacy warning.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The search-suggestions recipe explicitly directs a fallback search using the user's original keywords. These requests use https://www.aconfig.cn (L003), but the file does not warn that user search terms are disclosed to this intermediary rather than sent only to YouTube.

Content

No source excerpt is available for this finding.

The API workflows lack a user-facing warning that video identifiers and URLs are sent to a third-party service.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

These chain recipes instruct API calls rather than merely listing available tools. The configured base URL is aconfig.cn (L003), and the endpoints accept video identifiers, video URLs, or subtitle URLs, but the file never instructs the agent to disclose that these request details are shared with a service other than YouTube.

Content

No source excerpt is available for this finding.

The documented no-write guarantee contradicts mandatory persistent filesystem writes.

Low
Category
Excessive Agency
Confidence
91% confidence
Finding

L261 explicitly states that the skill does not perform writes, but L329–L332 create a directory under the user's home directory, overwrite a confirmation file, and append an audit log. Further mandatory audit writes occur at L354 and L431–L435, so execution has persistent local side effects despite the unqualified no-write guarantee.

Content

No source excerpt is available for this finding.

The standalone trigger “Shorts” does not establish intent to retrieve a channel’s Shorts list.

Low
Category
Trigger Abuse
Confidence
90% confidence
Finding

The recipe treats “Shorts” alone as an activation condition, although its workflow retrieves channel information and a channel-wide Shorts list. Requests about an individual Short or Shorts generally could therefore select this recipe without the required channel-level intent.

Content

No source excerpt is available for this finding.

Generic recommendation and trend phrases can activate video-specific recipes outside a video context.

Low
Category
Trigger Abuse
Confidence
94% confidence
Finding

L051 includes “推荐” (“recommend”), and L121 includes “趋势” (“trends”) as semantic activation phrases. Both occur commonly in unrelated requests, but the recipes do not explicitly require a video-related conversational context or exclude non-video requests; recipe titles and input requirements do not establish that activation boundary.

Content

No source excerpt is available for this finding.

Batch-operation triggers include everyday words without requiring a data-collection context.

Low
Category
Trigger Abuse
Confidence
94% confidence
Finding

Keywords such as “全部” (all), “所有” (every), and “循环” (loop) can occur in ordinary requests that do not involve bulk data collection. Although the category description mentions collection, the matching rules do not require that context or provide exclusions, so unrelated requests can trigger the bulk-collection confirmation.

Content

No source excerpt is available for this finding.

Confirmation rules require Chinese wording without offering a language choice.

Low
Category
Output Handling
Confidence
96% confidence
Finding

Every accepted confirmation is Chinese, while English responses such as “yes” are explicitly rejected. This requires users to confirm in a particular language without opt-in or a documented reason why equivalent explicit consent in their language is insufficient.

Content

No source excerpt is available for this finding.