ClawHub
Back to skill

Security audit

JSON修复工具

Security checks for vulnerabilities and agentic risk

Overview

This JSON repair skill is mostly purpose-aligned, but it asks for file-writing and command execution while giving inconsistent and overconfident safety guidance.

Review this skill before installing. Use it only on files you are willing to modify, prefer writing to a separate output file, avoid untrusted malicious JSON-like input, and do not rely on the vm sandbox wording as a strong security guarantee.

SkillSpector

Vulnerability Patterns
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (3)

Intent-Code Divergence

Medium
Confidence
93% confidence
Finding
The skill claims an 'Eval沙箱' using vm.runInNewContext as a safety control, but this markdown provides no substantiating implementation and elsewhere describes direct Node.js/CLI processing. Misstating sandboxing is dangerous because users may treat untrusted malformed input as safely isolated when Node vm contexts are not a robust security boundary and may not even be used at all.

Intent-Code Divergence

Medium
Confidence
84% confidence
Finding
The request format advertises a callback_url for asynchronous completion, but the FAQ later states the tool does not support asynchronous processing. This inconsistency can mislead integrators into enabling outbound callback behavior or building workflows around features that do not exist, increasing the risk of unsafe network assumptions, SSRF-style design mistakes, or insecure compensating logic.

Missing User Warnings

Medium
Confidence
88% confidence
Finding
The document describes file repair and output behavior, including overwrite-related backup options, without clearly foregrounding that files may be modified in place. This is dangerous because an agent with read/exec/write tools could alter user files unexpectedly, and the backup safeguard is optional and therefore not a complete mitigation.

Static analysis

No suspicious patterns detected.