Back to skill

Security audit

微博数据助手

Security checks for vulnerabilities and agentic risk

Overview

ClawScan’s verdict is benign with medium confidence: the scanners report no AIG findings, alongside concerns about undeclared compliance records, an unpinned installer, and some workflow and privacy issues; Skillspector coverage was partial.

The verdict is benign, but it does not mean risk-free. Before installing, consider the unpinned installer and the reported local compliance records, check that the consent and data-sharing practices suit you, and keep in mind that Skillspector marked its analysis partial because some local references were missing or ambiguous.

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 (13)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Local bookkeeping alone could be a supporting implementation detail, but this code explicitly asserts that legal confirmation occurred, including a timestamp and confirmed status, without collecting or checking consent. That is a substantive undeclared behavior, not merely an implementation detail of read-only Weibo analysis. This finding is limited to the supplied chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

The supplied code performs compliance bookkeeping rather than Weibo querying or analysis. Logging alone could be a supporting implementation detail, but this entry asserts a completed legal review unrelated to any observed query or review process. That compliance-state modification is not disclosed by the declared purpose. No API requests or data exfiltration are shown.

Content

No source excerpt is available for this finding.

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

Medium
Category
MCP Rug Pull
Confidence
92% confidence
Finding

The installation command invokes an unpinned npm package through npx, allowing the executed installer to change without a documentation change or review. A compromised upstream release could execute code with the installing user's privileges; although this is an installer rather than an MCP server, the supply-chain concern remains valid.

Content

No source excerpt is available for this finding.

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

Medium
Category
Output Handling
Confidence
96% confidence
Finding

The accepted confirmations are exclusively Chinese, while English responses such as 'yes' are explicitly rejected and users are directed to enter '同意'. This imposes a language requirement without user opt-in or a documented locale-specific justification, despite the manifest listing English support.

Content

No source excerpt is available for this finding.

The instructed API workflows lack a privacy warning that queried Weibo identifiers are sent to a third-party provider.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

L028–L035 instruct the agent to execute chained API requests, rather than merely listing available tools. These requests send user-provided or derived post/video identifiers to aconfig.cn (the base URL at L003), but the file does not warn users that this intermediary receives their lookup identifiers.

Content

No source excerpt is available for this finding.

Generic trigger keywords can activate Weibo recipes for unrelated requests.

Medium
Category
Trigger Abuse
Confidence
94% confidence
Finding

Triggers such as “看评论” (view comments), “看回复” (view replies), “高级搜索” (advanced search), and “收藏夹” (favorites) are common across many applications but route directly to Weibo recipes. The longest-match and tie-breaking rules do not require an established Weibo context, and no exclusions prevent activation for other platforms or general requests.

Content

No source excerpt is available for this finding.

The confirmation rules require Chinese consent wording without offering a language choice.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

Every accepted confirmation is Chinese, while English “yes” is explicitly rejected and the fallback prompt directs the user to enter Chinese wording. This imposes a language requirement on user consent without an opt-in or an explanation of why Chinese-only confirmation is necessary; the platform-specific context alone does not establish that necessity.

Content

No source excerpt is available for this finding.

L477–L482 authorize querying the MaxHub account balance, which is outside the stated public Weibo data-analysis purpose.

Low
Category
Excessive Agency
Confidence
91% confidence
Finding

The whitelist permits GET /api/maxhub/balance, exposing a service-account balance query rather than public Weibo content or analytics. The manifest does not declare account billing or balance management, and this file establishes no dependency that makes this capability necessary for the advertised workflows.

Content

No source excerpt is available for this finding.

The trigger “作者” (“author”) is too broad for a Weibo author-profile workflow.

Low
Category
Trigger Abuse
Confidence
96% confidence
Finding

“作者” can refer to the author of a book, article, or other content, not just a Weibo post. Although the recipe requires a Weibo status ID, the activation instruction does not explicitly require an established Weibo context or exclude unrelated author questions.

Content

No source excerpt is available for this finding.

The trigger “推荐” (“recommend”) can activate the Weibo feed workflow for unrelated recommendation requests.

Low
Category
Trigger Abuse
Confidence
98% confidence
Finding

“推荐” is common in requests for products, books, restaurants, and other suggestions. This trigger has no explicit Weibo-context restriction, and the recipe requires no input that would otherwise distinguish those requests from a request for the Weibo recommendation feed.

Content

No source excerpt is available for this finding.

The advanced-search workflow sends user-provided search terms and filters to an external search API without a privacy disclosure.

Low
Category
Output Handling
Confidence
88% confidence
Finding

This is an executable workflow step that maps the user's q and filter inputs into an API call, rather than merely listing an available tool. The recipe does not warn that these inputs will be transmitted to the external search service, which matters if a query contains sensitive information.

Content

No source excerpt is available for this finding.

The trigger “收藏夹” (“favorites/bookmarks”) is too broad for a user-video collection recipe.

Low
Category
Trigger Abuse
Confidence
96% confidence
Finding

“收藏夹” commonly refers to browser bookmarks and favorites in many applications, but this trigger selects a specific user-video collection workflow. The activation condition provides no platform qualifier or exclusions to distinguish those unrelated requests.

Content

No source excerpt is available for this finding.

Batch-collection warnings can be triggered by generic words unrelated to data collection.

Low
Category
Trigger Abuse
Confidence
94% confidence
Finding

The batch-operation keyword list includes everyday terms such as “全部” (all), “所有” (every), and “循环” (loop), without requiring a collection-related action or providing exclusions. Even within weibo-pro, requests such as explaining all available parameters can match these terms and unnecessarily trigger the batch-data-collection confirmation.

Content

No source excerpt is available for this finding.