Seedance 参考音频生视频

• Outcome: Review
• Latest audit: Aug 16, 2026 · 3:18 PM UTC
• Version: 1.0.0

Overview

The skill performs a disclosed AI Hive video workflow, but its activation wording is much broader than that workflow and could route unrelated tool-comparison or e-commerce requests into a credentialed media-upload service.

Install only if you specifically want AI Hive Seedance audio-reference video generation and are comfortable providing an AI Hive API key, uploading selected media to the service, and possibly incurring generation costs. Treat broad comparison, shopping-platform, or marketing-advice requests as informational unless you explicitly choose to generate a video.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep
• Trigger Abuse: Overly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
• MCP Least Privilege: Underdeclared Capability, Wildcard Permission, Missing Permission Declaration
• MCP Tool Poisoning: Hidden Instructions, Unicode Deception, Parameter Description Injection
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands

Findings (10)

Lp3

Medium

• Category: MCP Least Privilege
• Confidence: 86% confidence
• Finding: The skill documentation describes capabilities that read environment/config values, write local files, invoke shell commands, and perform network operations, yet no explicit permissions are declared. That creates a transparency and governance gap: users or orchestration systems cannot accurately constrain or review what the skill is allowed to do before execution. In this context the operations appear related to API initialization, upload, polling, and download rather than overtly malicious behavior, but the missing permission declaration still increases risk.

Tp4

High

• Category: MCP Tool Poisoning
• Confidence: 90% confidence
• Finding: The skill is presented as a narrowly scoped 'audio reference to video' tool, but the documented behavior includes broader model discovery, generic uploads, task handling, initialization flows, and even references to other generation/chat capabilities. Description-behavior mismatch is dangerous because it undermines informed consent and policy enforcement: a system may invoke the skill expecting limited audio-to-video behavior while it can access broader APIs and workflows. The wide gap is especially concerning because it includes credential setup and generalized media handling.

Description-Behavior Mismatch

Medium

• Confidence: 91% confidence
• Finding: The file implements a broad, general-purpose AI Hive client with chat, image, model listing, upload, and account-related functions, while the declared skill is narrowly scoped to Seedance audio-reference-to-video. In a skill ecosystem, this scope mismatch increases the attack surface and enables unintended capability exposure if the script is invoked outside the narrow wrapper path.

Context-Inappropriate Capability

Low

• Confidence: 80% confidence
• Finding: User-info retrieval is not required to generate a Seedance audio-reference-to-video result, yet the script exposes a command that can fetch account and wallet data. Even if not automatically invoked, bundling this unrelated capability in the skill unnecessarily expands access to sensitive account metadata.

Context-Inappropriate Capability

Medium

• Confidence: 93% confidence
• Finding: The generic text chat endpoint is unrelated to the skill's advertised purpose and allows arbitrary model interaction using the same API key and infrastructure. In a constrained skill environment, this creates an unjustified expansion from media generation to open-ended LLM access, which can be abused for data handling or policy bypass workflows outside the intended task.

Context-Inappropriate Capability

Medium

• Confidence: 92% confidence
• Finding: Image generation is a separate capability from the fixed Seedance audio-reference-to-video function but is implemented and exposed in the same script. This unnecessary capability broadening increases misuse potential and weakens assurances that the skill only performs the task users expect.

Intent-Code Divergence

Medium

• Confidence: 88% confidence
• Finding: The module-level documentation explicitly describes a general-purpose AI tool, contradicting the specialized Seedance skill declaration. This discrepancy is a security-relevant signal because it suggests the code was reused without proper reduction of unrelated features, making hidden or unintended behavior more likely.

Vague Triggers

Medium

• Confidence: 84% confidence
• Finding: The skill description contains an extremely broad set of trigger keywords spanning many AI video brands, e-commerce platforms, and generic creative intents. Overbroad invocation criteria can cause the agent to select this skill for loosely related user requests, leading to unintended external API calls, media uploads, or credential prompts when the user did not ask for this specific workflow. This is a routing/invocation safety issue rather than direct code execution, but it can still produce privacy, cost, and trust harms.

Vague Triggers

Medium

• Confidence: 85% confidence
• Finding: The markdown explicitly says the skill should be considered for many competitor, platform, pricing, migration, and API-search scenarios. That ambiguity expands the skill's reach beyond its core function and increases the chance of inappropriate invocation in comparative or informational queries that do not require uploads, task submission, or external service interaction. In context, this makes the skill more dangerous because it couples broad discovery terms with operational capabilities.

Vague Triggers

Medium

• Confidence: 83% confidence
• Finding: The e-commerce coverage section enumerates a very large set of marketplaces, ad products, merchant roles, and marketing objectives without clear limits. This can cause the agent to over-match ordinary commerce or marketing requests and invoke a skill capable of media upload and external task execution in situations where the user only wanted advice, comparison, or planning. The skill context does not justify such expansive routing boundaries, so the ambiguity meaningfully raises misuse risk.

VirusTotal

VirusTotal findings are pending for this skill version.

Static analysis

No suspicious patterns detected.