NVIDIA has released SkillSpector, an open-source scanner that flags vulnerabilities and malicious patterns in AI agent skills.
Summary
NVIDIA has released SkillSpector, an open-source scanner for agent skills that flags prompt injection, data exfiltration, unauthorized code execution, and malicious patterns in AI agent skills.
What changed
- SkillSpector scans agent skills — packaged capability bundles that help AI agents complete tasks — for security issues.
- It flags prompt injection, data exfiltration, unauthorized code execution, and malicious patterns.
- It inspects skills for hidden instructions, credential theft, and network activity.
- It can be used as a standalone scanner, or as part of a CI/CD pipeline.
- It is available as an open-source tool on GitHub, with a permissive Apache 2.0 license.
- It was developed by NVIDIA, in collaboration with researchers from the University of California, Berkeley, and the University of Illinois Urbana-Champaign.
- It was tested on 2,500+ agent skills, and found that 10% of them had at least one security issue.
- It found that 1 in 10 skills had at least one security issue, and 1 in 20 had a critical security issue.
- It found that the most common issues were prompt injection, data exfiltration, and unauthorized code execution.
- It found that the most common malicious patterns were hidden instructions, credential theft, and network activity.
Why it matters
- SkillSpector helps developers scan agent skills for security issues before they install them.
- It also shows how NVIDIA is working to improve the security of AI agent skills.
- The inclusion of a permissive Apache 2.0 license lets anyone use it, without a paid plan.
Sources
- SkillSpector — GitHub repository
- Scan Agent Skills Before Installation
- Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale
Originally published at IA Compa
Top comments (0)