NVIDIA has open-sourced a security tool called NVIDIA SkillSpector to scan artificial intelligence agent skills for vulnerabilities and malicious code before installation.
AI agent skills differ from standard text prompts because they bundle instructions with real executable code. Consequently, when an agent runs these skills, it operates with the same access permissions as the user. This level of access creates significant security risks for local environments.
Specifically, a compromised skill can read environment variables, extract API keys, and transmit sensitive data to unauthorized external servers. Recent research indicates that approximately 25 percent of public skills contain security vulnerabilities, while a smaller portion is intentionally malicious.
The Security Risks of AI Agent Skills
Developers often download skills to save time during development. However, the lack of security verification can expose local systems to severe exploits. Meanwhile, the rapid adoption of AI agents has outpaced the development of dedicated cybersecurity scanning tools.
How NVIDIA SkillSpector Analyzes Code
To address these vulnerabilities, NVIDIA SkillSpector allows users to analyze skills from multiple sources. Specifically, users can point the scanner at a local folder, a single markdown file, a GitHub repository link, or a compressed zip file.
Two-Pass Analysis and Risk Scoring
The scanning process consists of two distinct evaluation phases. First, a fast static analysis pass flags risky patterns such as credential harvesting, data leaks, and prompt injection. Furthermore, this initial pass checks code dependencies against live Common Vulnerabilities and Exposures (CVE) databases.
Second, an optional pass utilizes a large language model to analyze the developer’s intent and eliminate false positives. Ultimately, NVIDIA SkillSpector generates a risk score ranging from 0 to 100 alongside a clear verdict of safe, caution, or do not install.
Supported Platforms and Open Source Licensing
The tool is released under the Apache 2.0 open-source license, making it accessible for community modification. Currently, the software scans skills designed for Claude Code, Codex CLI, and Gemini platforms. As a result, developers can verify third-party apps and code before integration.
Future Outlook for AI Security
As autonomous agents become more common in enterprise workflows, securing executable skills will remain a priority. Notably, open-source tools like this scanner provide a baseline for verifying third-party code. Consequently, organizations can adopt AI automation with reduced exposure to digital threats.
Source: X (@akshay_pachaar)
