Back to skill

Security audit

今日头条数据助手

Security checks for vulnerabilities and agentic risk

Overview

ClawScan labels this skill benign with medium confidence: its results report no AIG findings and mostly describe content-query guidance, but they also flag undeclared local compliance records and an unpinned installer, while one scanner's review is partial.

The verdict is benign, but it does not mean risk-free. Before installing, consider whether you accept the reported local compliance and audit-log writes and the unpinned installer; scanner coverage was partial, so some local path references could not be resolved.

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
  • 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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (11)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied command performs software maintenance, not Toutiao data retrieval or analysis. Upgrading a skill is a distinct state-changing capability not disclosed in the description and is not merely an implementation detail of the declared querying behavior. The command alone does not establish which network endpoints it contacts.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

Ordinary operational logging could be a supporting implementation detail. This code instead creates a substantive compliance record claiming completion of a legal review, which is not disclosed by or necessary to the declared content-query purpose. The mismatch is the undeclared compliance-state change, not merely the absence of API calls in this chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

Ordinary local logging could support the declared skill without constituting a mismatch. This code goes further: it records a substantive legal-confirmation state as already confirmed, although the description does not disclose that behavior and the supplied chunk contains no consent step. The mismatch is the unconditional confirmation, not merely the absence of querying code or the presence of local file writes.

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 executes an unpinned npm package through npx, allowing the executable version to change without review. Although this is not an MCP server reference, a compromised future release could execute code with the installing user's permissions.

Content

No source excerpt is available for this finding.

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

Medium
Category
Output Handling
Confidence
98% confidence
Finding

The gate accepts only listed Chinese affirmations, explicitly rejects English “yes,” and stops all further operations until the required confirmation is supplied. This imposes a language requirement without user opt-in or a documented justification for restricting consent to Chinese.

Content

No source excerpt is available for this finding.

The manifest advertises search, recommendations, and trending-content queries, but the skill instructions explicitly prohibit all three.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The manifest states that the skill supports 搜索、热榜和推荐内容 and directs agents to match search and trending intents to recipes. In contrast, L039–L041 declare these capabilities unsupported, and L043 instructs the agent to refuse such requests, directly contradicting the advertised scope.

Content

No source excerpt is available for this finding.

The confirmation rules require Chinese-language assent without offering a language choice.

Medium
Category
Output Handling
Confidence
92% confidence
Finding

Every accepted confirmation is Chinese, while English “yes” and “ok” are explicitly rejected and the fallback asks for Chinese wording. This imposes a language-specific confirmation requirement without user opt-in or a documented justification for restricting the confirmation language.

Content

No source excerpt is available for this finding.

The skill declares read-only operation but requires persistent local filesystem writes.

Low
Category
Excessive Agency
Confidence
94% confidence
Finding

The manifest declares capability: read_only (L011), and the core constraints explicitly prohibit writes (L235). However, the mandatory compliance workflow creates a directory, overwrites a confirmation file, and appends an audit log (L302–L305); additional audit writes are required after successful API calls (L402–L406). These are limited compliance-related side effects rather than remote account mutations, but they contradict the unqualified read-only claim.

Content

No source excerpt is available for this finding.

The trigger “谁写的” (“who wrote it?”) can invoke an article-author recipe without establishing the relevant content or platform.

Low
Category
Trigger Abuse
Confidence
90% confidence
Finding

“谁写的” is a common authorship question that can refer to code, a message, a book, or other content unrelated to this recipe. The matching rules resolve competing triggers but do not require a supported article context before invocation.

Content

No source excerpt is available for this finding.

The trigger 「看看评论」 activates both article and video recipes without a disambiguation rule.

Low
Category
Trigger Abuse
Confidence
94% confidence
Finding

Both recipes explicitly list the same phrase, meaning “look at the comments,” as an activation condition. Neither specifies how to select the correct recipe when the request does not identify whether the target is an article or a video.

Content

No source excerpt is available for this finding.

Batch-operation reminders can be triggered by ordinary words unrelated to bulk collection.

Low
Category
Trigger Abuse
Confidence
95% confidence
Finding

The keyword list includes broad terms such as “全部” (all), “所有” (every), and “循环” (loop), without requiring a data-collection action or specifying exclusions. Within toutiao-pro requests, these words can occur in ordinary questions that do not request bulk collection, causing unintended compliance prompts.

Content

No source excerpt is available for this finding.