Back to skill

Security audit

西瓜视频数据助手

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because scanners report behavior beyond its stated read-only public-video purpose, including skill updates and local compliance records; the evidence is incomplete because some references could not be resolved.

Consider the Review verdict before installing. The scanners report undeclared updates and local records, a changeable installation package, and missing notices about information sent to aconfig.cn. Some local references could not be resolved, so the evidence is incomplete; the clean AIG scan does not rule out these reported concerns.

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 (15)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The error-handling guidance maps API failures to skill upgrade commands, including a forced upgrade, without explicitly requiring separate maintenance authorization or a reviewed target release. An agent could interpret this as permission to replace trusted skill content during a read-only query, potentially introducing attacker-controlled instructions if the update channel is compromised.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

This is more than incidental logging: the code unconditionally sets legal confirmation to 'confirmed' and records a first-time confirmation event. That consent-related state change is not declared by the public-video query and analysis description. Although the chunk may be part of a larger skill, its automatic confirmation behavior is independently an undeclared capability.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The only supplied executable behavior is an installation command, which is materially different from the declared read-only API operations and can modify the local environment. It also introduces package-resolution resources not covered by the stated request destination. The installed skill's behavior cannot be determined from this snippet; there is no evidence here that it implements the advertised queries.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

Ordinary operational logging could be a supporting implementation detail, but this code also unconditionally establishes affirmative legal-confirmation state and records that confirmation as an audit event. That is a substantive, undeclared state-changing action rather than merely supporting read-only data retrieval. No external requests or data exfiltration are shown in this chunk.

Content

No source excerpt is available for this finding.

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

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

The installation command uses npx without pinning the skillhub package version, allowing newly published package code to execute with the installing user's privileges. This is a supply-chain exposure rather than an MCP-server configuration issue; installation is legitimate, but the executable dependency is mutable.

Content

No source excerpt is available for this finding.

The compliance gate requires Chinese-language confirmation without offering an equivalent language choice.

Medium
Category
Output Handling
Confidence
97% confidence
Finding

The gate accepts only listed Chinese affirmative phrases and explicitly rejects English responses such as yes and ok. This forces users to confirm in Chinese regardless of their preferred language, without documenting a justification for the restriction.

Content

No source excerpt is available for this finding.

API failures can trigger skill upgrades despite the declared read-only query scope.

Medium
Category
Excessive Agency
Confidence
91% confidence
Finding

The upgrade guidance permits an interpretation in which ordinary API failures lead to software modification, including use of --force, without a distinct user-approval boundary. Although automatic execution is not demonstrated, following that interpretation could replace skill instructions through an unreviewed update during an otherwise read-only task.

Content

No source excerpt is available for this finding.

The search workflow lacks a warning that user-provided keywords are sent to a third-party API.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The workflow directs searches using user-provided keywords and chains the results into additional API calls. These requests use the aconfig.cn service identified at L003, but the file does not warn users that their search terms are disclosed to that service rather than only to Xigua.

Content

No source excerpt is available for this finding.

The trigger “看看评论” (“look at the comments”) can invoke a video-specific recipe without establishing video context.

Medium
Category
Trigger Abuse
Confidence
88% confidence
Finding

“看看评论” can refer to comments on articles, products, code, or other content, but it maps directly to video_detail_with_comments. The matching rules resolve competing triggers by length or clarification, without requiring video context or excluding non-video requests.

Content

No source excerpt is available for this finding.

The lookup workflow lacks a privacy notice about sending user identifiers to a third-party API.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

These chain recipes direct API calls that transmit the target user's user_id to aconfig.cn, the service identified on line 3. Although the base URL is listed, the file does not warn users that their lookup identifiers will be shared with this third-party service; these executable workflows go beyond merely listing available tools.

Content

No source excerpt is available for this finding.

The read-only skill mandates persistent writes to the user's home directory.

Low
Category
Excessive Agency
Confidence
97% confidence
Finding

The manifest declares a read-only capability, and L240 explicitly says the skill does not perform writes. However, L307–L310 create a directory, overwrite a confirmation file, and append an audit log under $HOME/.maxhub-skills/.compliance; additional log writes are required at L331, L346, and L411. These are persistent local side effects not disclosed by the read-only description.

Content

No source excerpt is available for this finding.

L60–L66 authorize querying the MaxHub account balance, which is outside the stated public-video analysis purpose.

Low
Category
Excessive Agency
Confidence
90% confidence
Finding

The whitelist permits GET /api/maxhub/balance, exposing account-level billing information rather than public Xigua video, author, comment, or discovery data. Balance retrieval is not necessary for the declared content-analysis workflows, and the manifest does not declare account or billing management.

Content

No source excerpt is available for this finding.

The unsupported-capability handler mandates a Chinese-language response without user language choice.

Low
Category
Output Handling
Confidence
91% confidence
Finding

The instruction requires the agent to reply with a fixed Chinese sentence whenever an unsupported capability is requested. It provides no language opt-in or instruction to adapt the response to the user's language, and the service-specific context does not establish a need for Chinese-only replies.

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
92% confidence
Finding

The keyword list includes broad terms such as “全部” (all), “所有” (every), and “循环” (loop). Although the category describes bulk data collection, no matching constraint or exclusion is specified to prevent ordinary requests containing these words from triggering a bulk-collection warning.

Content

No source excerpt is available for this finding.

The confirmation policy requires Chinese-language consent without offering a language choice.

Low
Category
Output Handling
Confidence
95% confidence
Finding

All accepted confirmations are Chinese, while the English affirmative “yes” is explicitly rejected and the fallback prompt directs users to enter “同意”. This imposes a language requirement on consent rather than merely documenting the skill in Chinese, without an explicit opt-in or justification.

Content

No source excerpt is available for this finding.