ClawHub
Back to skill

Security audit

Dlazy Generate

Security checks across malware telemetry and agentic risk

Overview

This skill appears to be a real dLazy media-generation wrapper, but it can run local CLI commands, upload prompts and local media to hosted services, and persist an API key with overly broad activation wording.

Install only if you intend to let an agent run the dLazy CLI for media generation. Avoid passing sensitive prompts or local files unless you are comfortable uploading them to dLazy-hosted services, prefer DLAZY_API_KEY for temporary use when possible, and review the exact CLI package/version before use.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (5)

Intent-Code Divergence

Medium
Confidence
99% confidence
Finding
The documentation is internally inconsistent: it states that no extra API key is needed, while earlier sections require a dLazy API key and describe how to store it. This can mislead users and agents into improper authentication handling, failed execution, or unsafe assumptions about where credentials are required and persisted.

Intent-Code Divergence

Low
Confidence
91% confidence
Finding
The skill minimizes its resource access by saying it does not access network or filesystem resources beyond the CLI, yet the same document explicitly states that prompts are sent to a hosted API and local files may be uploaded. This downplays real data egress behavior and can cause users to expose sensitive prompts or files without clear informed consent.

Vague Triggers

High
Confidence
95% confidence
Finding
The trigger keywords are extremely broad ('generate', 'create image, video, audio') and overlap with ordinary user requests, making accidental invocation likely. In a skill with exec capability that directs the agent to run shell commands and potentially upload local files, overbroad activation materially increases the chance of unintended command execution and external data transfer.

Missing User Warnings

Medium
Confidence
93% confidence
Finding
The skill instructs the agent to run `dlazy <model_name> -h` and then execute generation commands, but the user-facing description does not clearly warn that shell commands will be run on the local system. This reduces transparency and informed consent for an exec-enabled skill, increasing the risk of users unknowingly authorizing command execution and downstream network activity.

Session Persistence

Medium
Category
Rogue Agent
Content
## Trigger Keywords

* generate
* create image, video, audio
* multimodal generation

## Authentication
Confidence
81% confidence
Finding
create image, video, audio * multimodal generation ## Authentication All requests require a dLazy API key. The recommended way to authenticate is: ```bash This runs a device-code flow (also works i

VirusTotal

VirusTotal findings are pending for this skill version.

View on VirusTotal

Static analysis

No suspicious patterns detected.