Back to skill

Security audit

Jits Builder

Security checks across malware telemetry and agentic risk

Overview

The skill’s purpose is coherent, but it needs review because it can publish local services to public Cloudflare URLs with weak scoping and control.

Install only if you deliberately want generated mini-apps to be reachable through public Cloudflare tunnel URLs. Do not use it with secrets, private business data, internal tools, or unreviewed generated code. Prefer a hardened version that validates names and ports, verifies cloudflared from an official source outside /tmp, asks before public deployment, supports local-only serving, and automatically expires or cleans up running apps.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (8)

Intent-Code Divergence

Medium
Confidence
84% confidence
Finding
The script's advertised interface only mentions list and stop, but the implementation also includes hidden serve and tunnel subcommands that can publish content and expose a local service externally. Concealed functionality reduces user awareness and reviewability, making it easier to misuse the tool to host content or create unexpected network exposure.

Vague Triggers

Medium
Confidence
90% confidence
Finding
The usage examples are phrased as ordinary natural-language requests, which can cause an agent to invoke this skill on broad, everyday prompts rather than only when a user explicitly intends to generate and publish code. In this skill's context, unintended activation is more dangerous because the documented behavior includes generating an app and exposing it via a public Cloudflare tunnel, increasing the chance of accidental code generation and exposure.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The README prominently advertises that generated apps are deployed to a public trycloudflare.com URL but does not warn users that this exposes locally served, AI-generated code to the internet. In this context, that omission is significant because generated apps may contain insecure logic, secrets, debugging endpoints, or unsafe dependencies, and publishing them externally greatly expands attack surface.

Vague Triggers

Medium
Confidence
91% confidence
Finding
The skill can be invoked by very generic, everyday phrases like 'Build me a pomodoro timer' or 'Make a tip calculator,' which overlap with normal user requests and increase the chance of unintended activation. Because the skill does more than simple content generation—it writes files, serves them locally, and exposes them through a public tunnel—overbroad triggering materially raises the risk of accidental deployment of user-described content.

Missing User Warnings

High
Confidence
97% confidence
Finding
The skill describes automatic deployment via a Cloudflare tunnel but does not clearly warn users that generated apps become publicly reachable on the internet. This can lead users to unintentionally expose sensitive content, internal-only tools, or generated pages containing confidential data, especially since the workflow implies immediate publication after a casual request.

Missing User Warnings

Medium
Confidence
91% confidence
Finding
The tunnel command invokes cloudflared to expose a localhost service on a public trycloudflare.com URL without any authentication, access control, or prominent warning to the user. This can unintentionally publish local content or services to the internet, increasing the attack surface and potentially leaking sensitive data.

Vague Triggers

Medium
Confidence
86% confidence
Finding
The description advertises an extremely broad capability: users can 'describe what you need' and receive a 'working tool deployed in seconds,' which implies open-ended activation and autonomous app generation/deployment without clear constraints. In an agent skill context, this can encourage unsafe interpretation of arbitrary natural-language requests and unintended sensitive actions such as generating, exposing, or deploying code without adequate scope limits or approval gates.

Tool Parameter Abuse

High
Category
Tool Misuse
Content
# Also kill tunnel if exists
      tunnel_pid=$(cat "$JITS_DIR/$name.tunnel.pid" 2>/dev/null)
      [ -n "$tunnel_pid" ] && kill "$tunnel_pid" 2>/dev/null
      rm -f "$JITS_DIR/$name".*
      echo "🛑 Stopped $name"
    else
      echo "❌ App '$name' not found"
Confidence
94% confidence
Finding
rm -f "$JITS_DIR/

VirusTotal

62/62 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.