ClawHub
Back to skill

Security audit

Claude Code Framework

Security checks across malware telemetry and agentic risk

Overview

This skill is an agent execution framework that claims to enforce approvals, but its own code allows approval-required tool calls to continue after only a warning.

Install only for review or in a tightly sandboxed environment. Do not connect it to real shell, write/delete, browser, messaging, package-install, or network tools unless the approval path is fixed to fail closed and you understand what logging or memory data it keeps.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (3)

Intent-Code Divergence

High
Confidence
98% confidence
Finding
The documentation claims every task is permission-checked, but the implementation only logs a warning for APPROVE-risk tools in DEFAULT mode and then continues execution. This creates a dangerous trust gap: operators may rely on stated approval guarantees while write, network, or other sensitive actions can still run without confirmation.

Intent-Code Divergence

Low
Confidence
92% confidence
Finding
The glob-to-regex conversion does not anchor patterns containing '.' or wildcards, so matching is effectively substring-based rather than whole-string based. In a risk-classifier, this can cause overbroad or unintended rule matches, potentially misclassifying tool calls and either allowing risky operations under the wrong rule or applying the wrong approval/block behavior.

Missing User Warnings

High
Confidence
100% confidence
Finding
When a tool is classified as APPROVE, the code emits a warning in DEFAULT mode but does not block or wait for user consent before executing the tool. This effectively bypasses the approval boundary for commands such as write, delete, curl/wget, git push, package installs, message, and browser actions, enabling unauthorized state changes or data exfiltration.

VirusTotal

65/65 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.