NVIDIA/SkillSpector
SkillSpector
SkillSpector es un escáner de seguridad diseñado para detectar vulnerabilidades y patrones maliciosos en plugins y extensiones de agentes de IA antes de que sean instalados. Funciona como un guardrail en tiempo de ejecución que calcula puntuaciones de riesgo numéricas y asigna etiquetas de severidad para proporcionar recomendaciones de instalación o bloquear extensiones externas arriesgadas.
El proyecto se distingue por utilizar modelos de lenguaje para realizar análisis de código semántico, evaluando la intención y el contexto del código para reducir los falsos positivos. También emplea la supresión de problemas basada en huellas digitales para rastrear e ignorar riesgos previamente aceptados a través de ciclos de escaneo repetidos.
La herramienta cubre la seguridad de la cadena de suministro de software mediante el escaneo de dependencias contra bases de datos de seguridad públicas y admite la ingesta de activos de múltiples fuentes desde directorios locales, URLs remotas y repositorios. Proporciona informes de vulnerabilidad en múltiples formatos legibles por máquina y por humanos para su integración en pipelines de CI/CD.
Features
- AI Agent Vulnerability Scanners - Analyzes AI agent plugins and extensions to identify malicious patterns and security vulnerabilities before installation.
- LLM-Based Analysis - Uses language models to perform semantic code analysis and evaluate intent to reduce false positives in vulnerability detection.
- LLM-Powered Semantic Analyzers - Uses a language model to evaluate code semantics and reduce false positives during security analysis.
- Security Intent Analysis - Uses large language models to evaluate code intent and context, reducing false positives in security detection.
- CVE Dependency Scanners - Identifies known vulnerabilities in the skill dependency tree by querying public security databases.
- Dependency Vulnerability Scanners - Identifies known vulnerabilities in third-party dependencies by querying public security databases.
- Security Guardrails - Provides runtime guardrails that block the installation of risky AI skills based on security risk scores.
- AI Agent Scanners - Serves as a specialized security scanner for detecting vulnerabilities and malicious patterns in AI agent skills.
- Software Supply Chain Security - Secures the AI plugin supply chain by scanning external dependencies and remote repositories for known CVEs.
- Asset Risk Scoring - Calculates numeric risk scores and severity levels for AI plugins to provide installation recommendations.
- Vulnerability Dependency Mapping - Implements the process of mapping a plugin's dependency tree against public security databases to identify known vulnerabilities.
- Static Code Analysis - Performs static analysis of plugin source code using LLMs to detect security risks without executing the code.
- Fingerprint-Based Suppressions - Employs unique hashes to track and suppress previously accepted security risks across repeated scan cycles.
- Multi-Source Asset Ingestion - Retrieves target code from local directories, remote URLs, or repositories for unified security processing.
- File and URL Scans - Inspects files from remote URLs, repositories, and local directories to identify security risks in AI skills.
- AI Application Security - Security scanner for detecting malicious patterns in AI agent skills.
