Krux
March 18, 2026
Chainguard Lets You Harden AI Agents With One File
Published: March 18, 2026 at 12:38 AM
Updated: March 18, 2026 at 12:38 AM
100-word summary
Chainguard just opened beta access to Agent Skills, a catalog of pre-hardened tools for AI agents that you can drop in by copying a single SKILL.md file. No new toolchain required. The catalog blocks common attack vectors like unrestricted shell access and overly broad permissions. When upstream sources change or security rules update, every skill automatically re-hardens itself. The catalog shows active maintenance, with skills like plantuml-ascii and deploy-to-vercel updated through late February and March 2026. Translation: your AI agents can do useful work without handing them the keys to your entire codebase. Chainguard is betting that pre-vetted, scoped permissions will become table stakes for anyone shipping agents into production.
What happened
Chainguard just opened beta access to Agent Skills, a catalog of pre-hardened tools for AI agents that you can drop in by copying a single SKILL.md file. No new toolchain required. The catalog blocks common attack vectors like unrestricted shell access and overly broad permissions. When upstream sources change or security rules update, every skill automatically re-hardens itself. The catalog shows active maintenance, with skills like plantuml-ascii and deploy-to-vercel updated through late February and March 2026. Translation: your AI agents can do useful work without handing them the keys to your entire codebase.
Why it matters
Chainguard is betting that pre-vetted, scoped permissions will become table stakes for anyone shipping agents into production.