NVIDIA has launched a verified skills program for autonomous AI agents to address the growing need for capability governance and operational trust. As agents become more powerful and utilize open tools and portable skills, organizations require more than runtime guardrails to ensure safety. They need to understand the origins, risks, and modifications of the instructions agents execute. NVIDIA-verified skills fill this gap by providing transparency, provenance, security validation, and authenticity checks at the agent capability layer. These agent skills are portable instruction sets that teach AI agents how to correctly use NVIDIA CUDA-X libraries, AI Blueprints, and platform tools. Published in the NVIDIA/skills GitHub repository, verified skills are cataloged, scanned, signed, and documented with a machine-readable skill card. This system builds on the open skills specification, ensuring that a single SKILL.md file works reliably across various coding agents like Claude Code, Codex, and Cursor. The verification process begins in a product team's source repository and moves through a flow involving human review, automated policy checks, and security scanning. A critical component of this process is SkillSpector, a tool that treats skills as deployable capabilities rather than static prompts. It scans for conventional software risks like vulnerable dependencies and suspicious scripts, as well as agent-specific dangers such as prompt injection, hidden instructions, and tool poisoning. This scanning aligns with recognized security standards, including OWASP and MITRE ATLAS guidance, helping NVIDIA block or remediate risky skills before they enter the public catalog. To ensure integrity, NVIDIA is experimenting with cryptographic signing. This signature covers every file within a skill directory, allowing developers to verify that the downloaded content is authentic and unaltered. Unlike simple registry listings, this method provides verifiable proof that the asset has not been tampered with since publication. Trust is further centralized through the skill card, a structured metadata file that accompanies each verified skill. The card answers essential questions for developers and enterprise architects, such as who authored the skill, what resources it accesses, and whether it has been validated against real-world benchmarks. For example, a developer deploying a delivery-scheduling agent can review the skill card for the NVIDIA cuOpt routing skill to confirm its safety and dependencies without manual auditing. NVIDIA distinguishes these verified skills from standard assets by embedding trust into the workflow itself. While runtime controls govern agent behavior during execution, verified skills govern the capabilities entering the workflow, extending AI governance across coding tools and enterprise platforms. The company has also released a skill card template and generator to help the community adopt these transparency standards. Organizations deploying agents in real environments can now pull skills from the catalog, verify their cryptographic signatures, and review skill cards to confirm ownership, license, and verification status. This approach ensures that capabilities come from trusted sources, have passed rigorous security checks, and remain unchanged after publication, enabling developers to extend autonomous agents with confidence.