Security audit

AI大模型专家｜短剧调色与画面统一

Security checks across malware telemetry and agentic risk

Overview

The skill has a legitimate short-video editing workflow, but it bundles broader AI-HIVE media-generation tools and credentialed upload/download behavior that are not tightly scoped to the advertised color-consistency purpose.

Install only if you are comfortable sending selected media and prompts to AI-HIVE under your own API key. Use authorized, non-sensitive footage, review generated commands before running them, avoid saving production secrets in shared environments, and treat the bundled image/video generation scripts as broader than a narrow color-grading helper.

SkillSpector

By NVIDIA

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection

Findings (12)

Lp3

Medium

Category: MCP Least Privilege
Confidence: 89% confidence
Finding: The skill advertises executable scripts that can read environment variables, access local files, invoke shell tooling like ffmpeg, and communicate with a remote AI-HIVE service, yet no explicit permissions are declared. This creates a consent and review gap: users and platforms may underestimate the skill's effective capabilities, especially because it uploads media, initializes API keys, polls tasks, and downloads results.

Tp4

High

Category: MCP Tool Poisoning
Confidence: 95% confidence
Finding: The documented purpose is narrowly framed as short-drama color grading and visual consistency, but the behavior described includes broad AI-HIVE media generation, upload/download workflows, API-key setup, model operations, and shared chat capabilities, while lacking actual color-matching or skin-tone correction functionality. This mismatch can mislead users into granting trust, sensitive media, or API credentials to a much more general remote-content pipeline than expected.

Description-Behavior Mismatch

High

Confidence: 98% confidence
Finding: The embedded skill configuration and CLI entrypoint implement a GPT Image 2 image-generation/editing tool, while the published skill metadata claims a much narrower short-drama color-grading and visual-consistency purpose. This mismatch is dangerous because it can mislead reviewers and users about the real capabilities being exposed, increasing the chance that broader model access, arbitrary prompt use, and media processing are granted under false pretenses.

Description-Behavior Mismatch

Medium

Confidence: 89% confidence
Finding: The module docstring describes a general-purpose AI Hive client covering chat, image, video, model queries, uploads, and polling rather than a limited drama color-consistency utility. In context, this broad self-description reinforces that the code does more than the skill claims, undermining transparency and making permission/risk evaluation unreliable.

Context-Inappropriate Capability

Medium

Confidence: 95% confidence
Finding: The client exposes user-info, model enumeration, text chat, upload-token creation, and task inspection endpoints that are unrelated to the advertised short-drama color-grading workflow. Extra capabilities increase attack surface and can be abused for account reconnaissance, unintended data handling, or repurposing the skill beyond what users and reviewers expect.

Description-Behavior Mismatch

High

Confidence: 98% confidence
Finding: The skill is presented as a narrowly scoped short-drama color-grading/visual-consistency tool, but the implementation is a broad AI Hive client that supports generic chat, model enumeration, uploads, and media generation. This capability mismatch is dangerous because it gives the skill substantially more power than users and reviewers would reasonably expect, enabling off-scope data access and arbitrary remote API use under a misleading description.

Intent-Code Divergence

Medium

Confidence: 90% confidence
Finding: The module docstring explicitly describes a general-purpose AI capability wrapper, which contradicts the skill metadata's specialized color-grading/visual-consistency claim. This inconsistency undermines trust boundaries and informed consent, because operators may approve or invoke the skill assuming limited functionality while the code advertises and exposes much broader actions.

Context-Inappropriate Capability

Medium

Confidence: 92% confidence
Finding: The user-info endpoint retrieves account and wallet information that is unrelated to the claimed video color-consistency purpose. In a misleadingly scoped skill, this expands access to sensitive account data without a clear business need, increasing the risk of unnecessary exposure or misuse.

Context-Inappropriate Capability

High

Confidence: 97% confidence
Finding: The unrestricted text chat function allows arbitrary prompts and model interaction, which is far outside a specialized color-grading skill's expected behavior. In context, this broad remote execution surface can be used to process unrelated or sensitive content, bypass workflow constraints, and turn the skill into a concealed general-purpose AI gateway.

Context-Inappropriate Capability

Medium

Confidence: 89% confidence
Finding: The model discovery functionality exposes the full set of available AI Hive models, which exceeds the declared use case and helps users or downstream automation repurpose the skill for unrelated capabilities. In context, this is dangerous because it weakens scope controls and makes it easier to pivot from a niche media tool into a broader multi-model access client.

Missing User Warnings

Medium

Confidence: 87% confidence
Finding: The skill describes uploading materials to AI-HIVE and downloading generated results, but does not provide a prominent warning about external data transmission, retention, or handling of potentially sensitive media. In a media-editing workflow, source footage may contain unreleased content, personal data, branded assets, or licensed material, so silent transmission materially raises privacy and confidentiality risk.

Vague Triggers

Medium

Confidence: 92% confidence
Finding: The skill enables implicit invocation but provides no narrowly scoped trigger conditions, so the platform may activate it for broadly related requests. Because this skill covers wide creative/editing tasks and can influence generated commands and workflows, ambiguous routing can cause unintended use, prompt-scope confusion, or accidental handling of requests the user did not clearly delegate to this skill.

VirusTotal

64/64 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

Detected: suspicious.exposed_secret_literal

File appears to expose a hardcoded API secret or token.

Critical

Code: suspicious.exposed_secret_literal
Location: SKILL.md:131