Kling 3.0 — AI Video Generation by Kuaishou

• Outcome: Pass
• Latest audit: May 29, 2026, 8:16 PM
• Version:
  1.0.4

Overview

This is a disclosed Atlas Cloud video-generation skill that uses an API key and may upload prompts or media for paid video generation, with some scope-broadening users should notice.

Install only if you intend to use Atlas Cloud for AI video generation or editing. Use a dedicated API key, review the model, duration, media inputs, and expected cost before running commands, and avoid uploading private or regulated media unless Atlas Cloud's data handling is acceptable to you.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Trigger Abuse: Overly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
• MCP Tool Poisoning: Hidden Instructions, Unicode Deception, Parameter Description Injection
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access

Findings (4)

Tp4

High

• Category: MCP Tool Poisoning
• Confidence: 87% confidence
• Finding: The skill advertises itself narrowly as Kling 3.0 / O3 video generation, but the documented interface allows listing models, uploading local files, and potentially invoking broader Atlas Cloud model behavior than users would reasonably expect from the description. That mismatch can cause unintended data disclosure or unintended operations, especially when local media upload and broader model selection are available but not prominently disclosed in the top-level description.

Description-Behavior Mismatch

Medium

• Confidence: 90% confidence
• Finding: The script’s documented and implemented scope is broader than the manifest claims: it supports generic Atlas Cloud video generation, arbitrary models, and video-to-video workflows rather than a Kling-only capability. This is dangerous because users or orchestrators may invoke the skill under a narrower trust assumption while the code can send prompts and media to other models/services than expected, creating supply-chain, privacy, and policy-boundary risk.

Description-Behavior Mismatch

Low

• Confidence: 92% confidence
• Finding: The code exposes model enumeration and arbitrary user-selected model submission, which exceeds the stated Kling-only behavior. This increases risk because an attacker or unexpected caller can route data to any available Atlas Cloud video model, bypassing user expectations and potentially violating data-handling or allowlist assumptions tied to the skill description.

Vague Triggers

High

• Confidence: 91% confidence
• Finding: The trigger language is extremely broad and can capture generic requests about making videos, editing content, product demos, reels, and other common tasks. Over-broad auto-triggering increases the chance that unrelated user content, local media, or sensitive prompts are routed to a third-party API without sufficiently specific user intent or informed consent.

VirusTotal

64/64 vendors flagged this skill as clean.