Back to skill

Security audit

AI开发工程师全流程工作流

Security checks across malware telemetry and agentic risk

Overview

This skill is a broad AI application engineering workflow with expected templates and helper scripts; its network and logging examples are disclosed and purpose-aligned, but users should handle prompts, logs, and third-party APIs carefully.

Install if you want a broad AI development workflow, but treat it as engineering guidance rather than a privacy-complete implementation. Before running evals or adding moderation providers, review what prompts, user text, logs, and test data will leave your environment, add consent and redaction where needed, and use mock/local modes for confidential datasets.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (8)

Vague Triggers

Medium
Confidence
90% confidence
Finding
The trigger list is extremely broad and overlaps with ordinary AI, coding, architecture, safety, and deployment requests, making accidental activation likely. Over-broad invocation increases the chance that this skill overrides more specific skills or engages in powerful engineering workflows without the user intentionally selecting it.

Vague Triggers

Medium
Confidence
88% confidence
Finding
The stage-one activation condition includes a generic phrase equivalent to 'help me design an AI system technical solution,' which is common in normal conversation and not strongly scoped to this skill. That ambiguity can cause the skill to activate in contexts where users expected lightweight advice rather than a broad engineering workflow with code-generation and architecture consequences.

Vague Triggers

Medium
Confidence
92% confidence
Finding
The full-process mode activates on a very broad phrase like 'help me build an AI app from scratch,' which can encompass many benign requests and implicitly authorizes a large end-to-end workflow. In a skill that covers architecture, coding, deployment, monitoring, and safety, such broad activation expands the attack surface for unintended actions and prompt hijacking via normal-looking requests.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The guide explicitly recommends collecting user feedback, corrections, and conversation interruption signals as part of an LLMOps feedback loop, but it does not mention data minimization, consent, retention limits, access controls, or redaction of sensitive content. In an AI application context, these logs commonly contain prompts, personal data, proprietary business information, or secrets, so omitting privacy and secure-handling guidance can lead teams to deploy unsafe telemetry pipelines by default.

Missing User Warnings

Medium
Confidence
93% confidence
Finding
The document recommends sending raw user-generated text to third-party content moderation providers but does not pair that guidance with user notice, consent, data-minimization, or privacy-handling requirements. In a safety-guardrails manual, this omission is meaningful because the transmitted text may contain sensitive personal, confidential, or regulated data, creating compliance and privacy exposure.

Missing User Warnings

Medium
Confidence
89% confidence
Finding
In non-mock mode, the script sends test queries and optional system prompt content to the OpenAI API, which means evaluation data may leave the local environment and be processed by a third party. Because the CLI does not clearly warn users at runtime that test inputs may contain sensitive prompts, datasets, or proprietary examples, operators can unintentionally exfiltrate confidential data during evaluation.

External Transmission

Medium
Category
Data Exfiltration
Content
def aliyun_content_check(text: str) -> dict:
    """阿里云内容安全审核"""
    # 检测: 色情、暴恐、违禁、广告、辱骂、灌水
    response = requests.post(
        "https://green.cn-shanghai.aliyuncs.com/green/text/scan",
        json={
            "scenes": ["antispam"],
Confidence
90% confidence
Finding
requests.post( "https://

External Transmission

Medium
Category
Data Exfiltration
Content
def aliyun_content_check(text: str) -> dict:
    """阿里云内容安全审核"""
    # 检测: 色情、暴恐、违禁、广告、辱骂、灌水
    response = requests.post(
        "https://green.cn-shanghai.aliyuncs.com/green/text/scan",
        json={
            "scenes": ["antispam"],
Confidence
90% confidence
Finding
requests.post( "https://green.cn-shanghai.aliyuncs.com/green/text/scan", json=

VirusTotal

64/64 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.