Cloudsmith Extends Artifact Registry Reach to AI Skills

Cloudsmith this week at the KubeCon + CloudNativeCon North America 2026 conference unveiled a registry for artificial intelligence (AI) skills that makes it simpler for software engineers to expose a set of reusable instructions, workflows, and resources that teach an AI coding agent how to better perform a specific task.

James Matchett, head of security for Cloudsmith, said the AI Agent Skills registry provides access to a set of vetted skills to ensure that application development teams are not inadvertently using an AI skill that might include a set of malicious prompts.

While it’s not clear to what degree software engineering teams are making use of AI skills, a survey of 505 engineers in the U.S. and United Kingdom conducted by Cloudsmith finds 83% are relying on platforms for managing software artifacts that were not built for the AI era even though they already identify malicious dependencies being added to AI-generated code as the top threat to software supply chains.

The AI Agent Skills registry minimizes that risk by hosting skills, including SKILL.md and supporting files, next to other software artifacts. DevSecOps teams can then define policies that govern usage of the AI skills, including making use of allow and deny lists that block an AI skill by name and version. Cloudsmith also provides access to observability tools that can be used to track what AI skills are being accessed and used by software engineering teams.

Unfortunately, not every software engineering team is making use of a registry to manage what software artifacts developers are allowed to use. As a result, many developers wind up downloading software artifacts directly from, for example, public repositories such as GitHub. Cybercriminals can embed malicious instructions in AI skills, Matchett said. In a technique known as slopsquatting, attackers publish malicious packages under names invented by AI coding tools.

In general, malicious actors are now essentially extending the same tactics and techniques they have used to compromise software supply chains to AI artifacts. The Cloudsmith survey finds that while well over half of respondents (58%) are moderately confident in the ability of their existing platforms to prevent a breach, only 37% can automatically identify, block, and trace the blast radius of a compromised artifact within minutes.

Eventually, most software engineering teams will need to revisit how they implement DevSecOps best practices for the AI era. As teams build and update applications faster, a single compromised package can now potentially infect hundreds of applications before being discovered, noted Matchett. DevOps teams need to strike a balance between the pace of application development and the risk a poisoned package or malicious artifact poses to the business, he added.

At this juncture, some type of breach of a software supply chain involving AI artifacts is all but inevitable for most organizations. The issue now is finding the best way to first minimize the number of incidents that might occur and, just as importantly, containing the breach as quickly as possible.