Back to skill

Security audit

Desktop Gui

Security checks across malware telemetry and agentic risk

Overview

This skill is a desktop automation guide, but its recommended vision workflow can upload full-screen screenshots to a hard-coded HTTP model endpoint without adequate privacy controls.

Install only if you intentionally want an agent to control your desktop. Use a VM or disposable desktop session, close sensitive windows, prefer local OCR or a local vision model, and avoid the remote vision example unless you control the endpoint, use HTTPS, crop/redact screenshots, and require manual approval before clicks or typing.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (6)

Description-Behavior Mismatch

Medium
Confidence
94% confidence
Finding
The documentation materially expands the skill from local GUI automation into a workflow that captures full-screen images and sends them to a remote vision API. That changes the trust boundary and can expose sensitive on-screen data such as credentials, messages, internal documents, or customer information to an external service without clear justification or containment.

Context-Inappropriate Capability

High
Confidence
98% confidence
Finding
The example shows arbitrary screenshot exfiltration to an external service that is not required for basic desktop automation. Screenshots of a real desktop can contain highly sensitive information, and transmitting them off-host creates a direct privacy and data-leak risk, especially when the feature is presented as a recommended mode.

Intent-Code Divergence

Medium
Confidence
88% confidence
Finding
The skill metadata and description present the capability as local tooling based on Python, xdotool, and scrot, but the recommended implementation depends on an undeclared remote Qwen vision service. This omission can mislead operators about network behavior and data exposure, undermining informed consent and security review.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The documentation instructs users to capture screenshots and send them to a remote model but does not clearly warn that desktop screenshots may contain confidential or regulated data. In a desktop-automation context, this omission is especially dangerous because the entire visible desktop may be exposed unintentionally.

External Transmission

Medium
Category
Data Exfiltration
Content
用 JSON 格式:{"action": "click", "x": 450, "y": 320, "description": "打卡按钮"}
"""

response = requests.post(
    "http://10.6.207.56:8000/v1/chat/completions",
    headers={"Authorization": "Bearer VLLM_API_KEY"},
    json={
Confidence
99% confidence
Finding
requests.post( "http://

External Transmission

Medium
Category
Data Exfiltration
Content
用 JSON 格式:{"action": "click", "x": 450, "y": 320, "description": "打卡按钮"}
"""

response = requests.post(
    "http://10.6.207.56:8000/v1/chat/completions",
    headers={"Authorization": "Bearer VLLM_API_KEY"},
    json={
Confidence
99% confidence
Finding
requests.post( "http://10.6.207.56:8000/v1/chat/completions", headers={"Authorization": "Bearer VLLM_API_KEY"}, json=

VirusTotal

65/65 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.