ClawHub
Back to skill

Security audit

Seedance 2.5 Reference to Video — 1080p Reference-Guided Video on RunComfy

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed RunComfy video-generation helper with some install and documentation caveats, but no hidden or purpose-mismatched behavior in the artifact.

Install only if you are comfortable sending your prompt and explicitly chosen reference media URLs to RunComfy for processing. Prefer a pinned or reviewed RunComfy CLI version instead of ad hoc latest npx use, protect the RUNCOMFY_TOKEN, and avoid using private media URLs unless they are intended for RunComfy to fetch.

SkillSpector

Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • 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 (6)

Vague Triggers

Medium
Confidence
94% confidence
Finding
The manifest says the skill triggers on "reference to video" and on "any explicit ask to generate video from reference images," which are broad activation conditions rather than narrowly scoped invocation phrases. This could cause unintended invocation when a user is discussing reference-based video generation generally, not specifically asking for this RunComfy Seedance skill.

Missing User Warnings

Medium
Confidence
91% confidence
Finding
This markdown file describes using the RunComfy-hosted model and later explains in the Security & Privacy section that prompts and reference URLs are transmitted to RunComfy. However, the main usage description does not prominently warn users near invocation or prerequisites that their media URLs and prompt data will be sent off-system to a third-party service.

Rp1

Medium
Category
MCP Rug Pull
Confidence
84% confidence
Finding
The skill recommends executing `npx -y @runcomfy/cli` without a pinned version, which allows the latest published package to be fetched and executed at run time. If the package is compromised, typo-squatted, or a malicious update is published, users may execute unreviewed code with their local privileges.

Rp1

Medium
Category
MCP Rug Pull
Confidence
84% confidence
Finding
The skill recommends executing `npx -y @runcomfy/cli` without a pinned version, which allows the latest published package to be fetched and executed at run time. If the package is compromised, typo-squatted, or a malicious update is published, users may execute unreviewed code with their local privileges.

Natural-Language Policy Violations

Low
Confidence
73% confidence
Finding
The documentation states that prompts accept Chinese or English, which presents a language constraint in natural language. Because no user choice, opt-in, or region-specific justification is provided, this can be read as a locale/language policy limitation.

Intent-Code Divergence

Low
Confidence
96% confidence
Finding
The documentation says at L174 that a reference video is mandatory on the 1080p endpoint, but elsewhere it explicitly states the opposite: L084 says the endpoint accepts requests with no `videos` key, and L185 labels `videos` as optional. This is an active contradiction in the skill's own documentation about core behavior.

Static analysis

No suspicious patterns detected.