Security audit

Dlazy Kling Audio Clone

Security checks across malware telemetry and agentic risk

Overview

This voice-cloning skill is purpose-aligned but needs Review because it uploads reference audio to a third-party service while lacking consent safeguards and containing inconsistent command/output documentation.

Install only if you intend to use dLazy's cloud voice-cloning service, have rights or consent for any voice sample you upload, and are comfortable storing a dLazy API key locally. Expect the current docs to be unreliable around examples and output type; use dlazy kling-audio-clone -h and verify arguments before running it.

SkillSpector

By NVIDIA

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
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
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access

Findings (7)

Intent-Code Divergence

Medium

Confidence: 93% confidence
Finding: The documented output schema claims this voice-cloning skill returns an image (`type: image`, `result.png`) instead of audio. This inconsistency can cause downstream agents or users to mis-handle generated content, weaken validation logic, and mask what data is actually being produced or transferred in a sensitive voice-cloning workflow.

Intent-Code Divergence

Medium

Confidence: 95% confidence
Finding: The usage examples and error table refer to a required `--prompt` parameter, while the help text only exposes `--audio_url` and `--name`. This mismatch can cause agents to invoke the tool incorrectly, produce unexpected failures, or route unintended user-controlled values into the wrong parameters, which is risky for an automated skill that uploads local files and calls remote APIs.

Intent-Code Divergence

Medium

Confidence: 92% confidence
Finding: The skill claims to perform audio cloning, but its documented output schema shows an image/PNG result. This mismatch can cause an agent or user to mis-handle returned artifacts, trust incorrect assumptions about file types, or route sensitive audio-processing tasks into the wrong downstream logic. In a skill that uploads local media and returns hosted outputs, inaccurate I/O documentation increases the chance of unsafe automation and data handling mistakes.

Intent-Code Divergence

Medium

Confidence: 90% confidence
Finding: The command examples use an undeclared --prompt argument even though the documented options only mention --audio_url and --name. This inconsistency can mislead agents into invoking unsupported parameters, causing failures or accidental misuse of a neighboring command path if the CLI accepts generic prompt-based flows elsewhere. Because this skill is meant to be executed by an agent, contradictory invocation guidance is operationally dangerous.

Vague Triggers

Medium

Confidence: 82% confidence
Finding: The trigger keywords are broad generic phrases such as '克隆声音' and '定制语音', which may overlap with ordinary user requests and cause accidental invocation. In a voice-cloning context, false activation is more dangerous than for a low-risk skill because it may prompt collection or upload of sensitive audio and initiate impersonation-capable generation without clear user intent.

Missing User Warnings

Medium

Confidence: 90% confidence
Finding: The skill description presents voice cloning as a straightforward customization feature but omits any warning about consent, impersonation, fraud, or legal/privacy risks. In this context, lack of warning materially increases the chance of misuse against third parties, especially because the workflow explicitly uploads reference audio to a hosted service for high-fidelity timbre replication.

Vague Triggers

Medium

Confidence: 83% confidence
Finding: Broad trigger phrases like 'clone voice' and 'custom speech' can cause the skill to activate in contexts beyond the narrow intended operation. Overbroad invocation surfaces increase the chance that an agent routes unrelated requests, potentially sending local audio or sensitive prompts to the external dLazy service without sufficiently specific user intent. Given that this skill uploads media to third-party endpoints, loose triggering is more risky than in an offline-only skill.

VirusTotal

62/62 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.