Back to skill

Security audit

video-stt

Security checks across malware telemetry and agentic risk

Overview

The skill does transcribe videos, but it needs review because normal use can change your machine and crafted command options could run unintended local code.

Install only if you are comfortable reviewing and running local scripts that download media, create transcript/audio files, create a Python virtual environment, and may install Homebrew or Python packages. Prefer installing dependencies yourself from trusted sources, avoid passing untrusted option values to the shell wrapper, and do not provide cloud API keys because this version does not actually implement cloud transcription.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (8)

subprocess module call

Medium
Category
Dangerous Code Execution
Content
result = subprocess.run(["which", cmd], capture_output=True)
        if result.returncode != 0:
            print(f"Warning: {cmd} not found. Installing...")
            subprocess.run(["brew", "install", cmd], check=False)

# 下载音频
def download_audio(url: str, output_dir: Path) -> str:
Confidence
95% confidence
Finding
subprocess.run(["brew", "install", cmd], check=False)

subprocess module call

Medium
Category
Dangerous Code Execution
Content
import whisper
    except ImportError:
        print("Installing whisper...")
        subprocess.run(["uv", "pip", "install", "whisper"], check=True)
        import whisper
    
    print(f"Loading Whisper {model} model...")
Confidence
96% confidence
Finding
subprocess.run(["uv", "pip", "install", "whisper"], check=True)

Lp3

Medium
Category
MCP Least Privilege
Confidence
89% confidence
Finding
The skill documentation clearly instructs users to run shell commands, create environments, install packages, and write output files, yet no corresponding permissions are declared. This creates a transparency and consent problem: an agent or user may invoke filesystem and shell-capable behavior without an explicit permission boundary, increasing the risk of unexpected local changes or abuse if companion scripts are unsafe.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding
The documented behavior exceeds the stated purpose by including environment setup, package installation, and local file writes, while also advertising cloud transcription modes that may not actually exist. This mismatch is dangerous because users may consent to transcription but not to system modification, dependency installation, or potential data transfer, leading to trust violations and unintended execution of high-impact operations.

Context-Inappropriate Capability

Medium
Confidence
98% confidence
Finding
A video transcription skill should not silently install system packages as part of normal execution. This expands the trust boundary from transcription into package management, enabling unexpected host changes and exposure to external package sources.

Context-Inappropriate Capability

Medium
Confidence
98% confidence
Finding
Installing Whisper on-demand during transcription adds host-modifying behavior that is not necessary for a runtime task and exposes the system to supply-chain and reproducibility risks. Users invoking transcription may not expect network access or package installation side effects.

Missing User Warnings

Medium
Confidence
92% confidence
Finding
The documentation does not warn that supplied URLs will be fetched/downloaded and that audio content may be transmitted to third-party transcription providers in cloud mode. This omission can expose sensitive media, internal URLs, or personal content to external services without informed consent, which is particularly risky for a skill designed around arbitrary user-provided video URLs.

Missing User Warnings

Medium
Confidence
97% confidence
Finding
Automatically modifying the system without explicit user confirmation is dangerous because it performs privileged or environment-altering actions unexpectedly. In an agent skill context, hidden installation behavior is more concerning because users may only intend to transcribe media, not authorize package management changes.

VirusTotal

66/66 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.