NVIDIA has released SkillSpector, an open-source security scanner designed to mitigate supply-chain risks in the rapidly expanding ecosystem of artificial intelligence agent skills. Introduced as part of the NVIDIA-Verified Agent Skills program in May 2026, the tool addresses a critical gap in AI security: the inability of traditional static analysis and code-scanning utilities to detect malicious instructions hidden within natural-language skill files. As organizations increasingly adopt AI agents, the proliferation of third-party skills has introduced severe vulnerabilities. Recent industry analyses indicate that over twenty-six percent of public skills contain prompt injections, while independent security audits have identified coordinated malware campaigns distributing skills designed to exfiltrate credentials and bypass safety guardrails. SkillSpector mitigates these threats by evaluating SKILL.md artifacts before installation. The scanner employs a two-tier architecture that combines deterministic static analysis with large language model-driven semantic evaluation. The initial deterministic layer inspects bundled scripts, tracks data flows, matches known vulnerability patterns, and cross-references open security databases. This stage is designed for speed and reproducibility. The subsequent semantic layer deploys multiple specialized LLM passes to identify paraphrased prompt injections, verify developer intent against declared permissions, and assess policy compliance. A final meta-analyzer adjudicates all findings, applying a confidence-weighted scoring algorithm that assigns point values based on severity and penalizes high-risk executable content. The system outputs a numeric rating from zero to one hundred, accompanied by a distribution recommendation. Independent testing demonstrates the necessity of this dual approach. When evaluated against real-world automation skills, the static layer frequently generates high volumes of benign findings, particularly when legitimate API calls or network requests match detection patterns. In one test case involving a standard GitHub automation skill, the static scan produced twenty medium-severity flags, all stemming from routine external transmissions. The semantic analysis successfully filtered these false positives while identifying four substantive risks, including autonomous code modification and persistent state manipulation. Crucially, the scanner proved capable of detecting prose-based prompt injections that traditional code scanners, repository vulnerability checkers, and credential hunters routinely miss. A deliberately constructed malicious fixture scored a perfect hundred, with its primary threat vector consisting entirely of natural-language instructions rather than executable code. Security researchers and platform engineers emphasize that the numeric score should not serve as an automated approval gate. The scoring model is intentionally calibrated to cast a wide net initially, relying on the semantic stage to refine precision. Consequently, human review remains essential to distinguish between aggressive feature implementation and genuine security threats. To facilitate integration into development workflows, SkillSpector can operate as a standalone command-line tool or function as a Model Context Protocol server, enabling automated installation gating within agent runtimes. The project is released under an Apache license and actively maintained, reflecting growing industry demand for specialized tooling that secures AI agent supply chains before they enter production environments.