SkillSpector - NVIDIA's open-source AI Agent skills security scanning tool - AiBoss

What is SkillSpector?

SkillSpector is an open-source AI Agent skills security scanning tool from NVIDIA. It automatically detects vulnerabilities, malicious patterns, and security risks before skills are installed on platforms such as Claude Code, Codex CLI, and Gemini CLI. The tool supports various input formats, including Git repositories, URLs, compressed files, directories, and single files. It includes 64 vulnerability patterns across 16 security categories and provides a two-stage analysis pipeline: fast static scanning and optional LLM semantic evaluation. It outputs reports in terminal, JSON, Markdown, and SARIF formats, along with a risk score from 0-100 and remediation suggestions.

SkillSpector's main functions

Multi-format input scanningSupports Git repositories, URLs, compressed files, directories, and single files.
64 vulnerability patterns covering 16 security categoriesThis includes issues such as injection of prompts, data leakage, privilege escalation, supply chain manipulation, excessive proxying, output processing, system prompt leakage, memory poisoning, tool abuse, rogue proxies, trigger abuse, dangerous AST code, taint tracking, YARA signing, MCP minimum privileges, and MCP tool poisoning.
Two-stage analysis pipelineThe first stage is a fast static scan, and the second stage is an optional LLM semantic evaluation.
Real-time vulnerability lookupKnown CVEs can be queried via the OSV.dev API, and automatic offline rollback is supported.
Multi-format outputSupports terminal, JSON, Markdown and SARIF report formats.
Risk scoring systemRating: 0-100 points, with a severity level label and specific repair recommendations.

SkillSpector's technical principles

Phase 1 Static AnalysisIt performs fast pattern matching based on regular expressions and ASTs to detect dangerous code calls such as exec, eval, and subprocess. It also queries dependency vulnerabilities in real time through the OSV.dev API, scans all files, and has a high recall rate and a moderate precision rate.
Second-stage LLM semantic analysis (optional)It assesses context and intent through compatible endpoints such as OpenAI, Anthropic, or NVIDIA, filters false positives, provides human-readable explanations, and improves accuracy to approximately 87%. The LLM hints include anti-jailbreak protection mechanisms to prevent malicious manipulation of analysis results.
Risk score calculationCRITICAL problem: +50 points, HIGH: +25 points, MEDIUM: +10 points, LOW: +5 points. Executable scripts are given an additional 1.3 multiplier, and are ultimately graded as 0-20 (LOW/SAFE), 21-50 (MEDIUM/CAUTION), 51-80 (HIGH/DO NOT INSTALL), and 81-100 (CRITICAL/DO NOT INSTALL).

How to use SkillSpector

Environmental preparationClone a GitHub repository and create an active Python virtual environment, supporting either uv or pip.
Installation and Deployment:implement make install Command to install production dependencies, or make install-dev Install development dependencies.
Perform a scan:use skillspector scan The command is followed by the target path.
Configure LLM (optional):set up SKILLSPECTOR_PROVIDER Choose the OpenAI, Anthropic, or NVIDIA provider for environment variables, and configure the corresponding API key to enable semantic analysis.
View report:pass --format The parameter specifies the terminal, JSON, Markdown, or SARIF output format.--output Specify the path to save the report.

SkillSpector's core advantages

Agent Scenario SpecializationWe focus on security scanning of AI Agent skills, and have deeply optimized detection rules for platforms such as Claude Code, Codex CLI, and Gemini CLI.
Two-stage precision balancingStatic scanning enables rapid response, while LLM semantic analysis improves accuracy to approximately 87%, balancing efficiency and accuracy.
End-to-end risk coverageIt features 64 built-in vulnerability patterns across 16 security categories, enabling end-to-end detection from vulnerability injection to supply chain poisoning.
Real-time vulnerability lookupIt connects to the OSV.dev database to query known CVEs in real time without requiring an API key, and automatically rolls back offline to ensure the availability of the isolated environment.
CI/CD native integrationSupports SARIF standard format output, which can be directly integrated with GitHub Code Scanning and enterprise continuous integration pipelines.
Multi-format input compatibilitySupports scanning of skill packages in any form, including Git repositories, URLs, compressed files, directories, and single files, lowering the barrier to entry.

SkillSpector's project address

GitHub repositoryhttps://github.com/NVIDIA/skillspector

SkillSpector's Competitive Comparison

Comparison Dimensions	SkillSpector	Cisco Skill Scanner
Developer	NVIDIA	Cisco
Open source license	Apache 2.0 (completely open source)	Not open source (commercial product)
Detection Engine	Two-stage pipeline: Fast static analysis + optional LLM semantic evaluation	Four-engine layering: Static analysis + Behavioral data flow analysis + LLM semantic analysis + Cloud threat scanning
Coverage Platform	Claude Code, Codex CLI, Gemini CLI	OpenAI Codex, Cursor
Vulnerability Mode	64 vulnerability patterns, spanning 16 security categories	The specific number was not disclosed, but the emphasis was placed on multi-engine layered coverage.
Risk Score	A quantitative score of 0-100, with four severity levels (LOW/MEDIUM/HIGH/CRITICAL).	Severity level labels, combined with cloud threat intelligence context
Real-time vulnerability database	Connect to OSV.dev for real-time querying of known CVEs, no API key required, and supports automatic offline rollback.	Integrating Cisco cloud threat intelligence, relying on cloud-based queries
Output format	Terminal, JSON, Markdown, SARIF	SARIF, JSON
CI/CD Integration	Native support for SARIF, compatible with GitHub Code Scanning and general pipelines.	Official integration with GitHub Actions is provided, deeply adapting to DevSecOps workflows.
Analysis depth	Static AST analysis + LLM semantic interpretation (accuracy approximately 87%)	Behavioral data flow tracing + execution path simulation + cloud-based correlation analysis
Deployment method	A local CLI tool that supports offline static scanning.	The cloud-based hybrid architecture relies on a cloud scanning engine for some of its capabilities.

Application scenarios of SkillSpector

Developer pre-installation reviewPerform a quick scan before installing third-party AI skills locally, and decide whether to install or reject them based on the risk score.
Enterprise CI/CD security access controlIntegrate scanning into the continuous integration pipeline to automatically intercept high-risk skill submissions and achieve left-shift security control.
Skills Market Platform ReviewAI skill stores or MCP marketplaces perform automated security reviews before skills are listed and display trust scores to users.
Independent developer self-auditSkill creators proactively scan their own projects before publishing to identify and fix potential vulnerabilities, thereby enhancing the credibility of their work.
Security research batch analysisResearchers performed batch scans on large-scale skills datasets to analyze vulnerability distribution, malicious pattern trends, and ecosystem security baselines.