ClawHub
Back to skill

Security audit

Youtube Editor

Security checks across malware telemetry and agentic risk

Overview

This skill is a disclosed video-editing automation tool with normal local-processing and API-use risks, but no artifact-backed evidence of hidden, destructive, or deceptive behavior.

Install only if you are comfortable sending video audio and transcript-derived text to OpenAI and running local media/browser tooling. Review nano-banana-pro separately before using AI thumbnail generation, use limited or revocable API keys, and avoid untrusted avatar file paths until the local HTML escaping issue is fixed.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • 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 (2)

subprocess module call

Medium
Category
Dangerous Code Execution
Content
try:
        # SECURITY: Timeout prevents runaway processes
        subprocess.run(cmd, check=True, timeout=900)
        return True
    except subprocess.CalledProcessError:
        print("⚠️ Image generation failed.")
Confidence
79% confidence
Finding
This code executes an external script from a user-writable workspace location if present, effectively trusting another skill file without integrity verification. An attacker who can place or modify that script can obtain arbitrary code execution when this skill runs, and the embedded text claiming the integration is safe should increase suspicion rather than reduce it.

Ssd 1

Medium
Confidence
88% confidence
Finding
The transcript is inserted directly into the LLM prompt, so adversarial spoken or embedded text in the video can steer the model away from the intended metadata-generation task. In this script the impact is bounded because the model output is later used mainly as text and a simple character-action prompt, but it can still corrupt outputs, produce abusive content, or manipulate downstream image generation behavior.

VirusTotal

60/60 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.