nvidia/skillspector is a security scanner that finds vulnerabilities and malicious patterns in AI agent skills before they are installed. It detects prompt injection (a technique that injects instructions to change an agent's behavior), data exfiltration, and supply-chain risks together in Claude Code, Codex, and MCP skills. The latest release in the repository is v2.13.0, and the provided sources contain no changelog for this version, so what was added or changed cannot be confirmed.
This scanner runs as part of the NVIDIA Verified Skills pipeline, scanning, evaluating, and signing agent skills before publication. In effect, it provides an analysis pipeline that assesses risk before skills are installed and used. Scan targets are Git repositories, URLs, ZIP files, directories, and single files, and it offers static analysis, optional LLM semantic analysis, vulnerability lookups through OSV.dev, risk scores, and reports in several formats. The implementation language is Python, the license is Apache-2.0, and the repository has 20766 stars. The repository was created on March 21, 2026.
That said, because LLM semantic analysis is optional, the provided material does not say how far static analysis alone can catch risks. The criteria for the risk score and its thresholds are not given either. The only score on record is a momentum of 8.3 out of 10, and no other metrics were produced. It can be used as a first-pass screening step before installation, but since the decision criteria are not published, it is still too early to treat the results as a basis for trust.