AI Security
Chainguard expands AI agent security with hardened skills registry and hardening service
The software supply chain security company launches a public registry of over 1,000 hardened agent skills, internal skill management, and an automated hardening service to secure the fast-growing AI agent ecosystem.
The New Stack
Chainguard is doubling down on securing AI coding agents with a major expansion of its Agent Skills platform—moving beyond a simple catalog toward continuous hardening and governance.
What's New#
The company now offers three tiers of security:
- Public Registry: Over 1,000 hardened community agent skills, updated weekly
- Private Registry: A namespace-based system for organizations to store and version internal skills (skills.cgr.dev/<org>/<skill_name>:<version>)
- Hardening as a Service: A closed beta allowing teams to submit internal skills for automated review, remediation, and continuous hardening
As Dustin Kirkland, Chainguard's SVP of engineering, explained, the goal is to let teams "plug agents directly into their software build and review pipeline without worrying that a compromised skill might introduce vulnerabilities or exfiltrate data."
How the Hardening Works#
Rather than a one-time scan, Chainguard treats hardening as an ongoing process. The system scans skills against a ruleset that catches:
- Over-permissioned scopes and capabilities
- Obfuscated commands and base64 execution
- Credential harvesting behavior
- Downloads from untrusted or suspicious domains
When issues are found, AI automatically rewrites and hardens the skill. Each hardened skill ships with a HARDENING.md audit log showing what was detected and changed. Crucially, when upstream skills update or new attack patterns emerge, previously hardened skills are automatically re-evaluated—ensuring teams always pull current versions, not stale scans.
The Broader Vision#
Chainguard positions Agent Skills as an evolution of its earlier work securing containers and open source packages. The company sees the same pattern repeating: a new artifact class (agent skills) arrives, adoption accelerates, governance lags, and attack surfaces expand. Internal skills, meanwhile, are often scattered across Slack threads and ad-hoc folders with no versioning or access control—a practice Chainguard calls inadequate.
The private registry brings structure: discoverability across teams, version pinning, rollback capabilities, and strict entitlements scoped to organization namespaces. The beta also integrates Model Context Protocol (MCP) to connect hardening directly to how skills are exposed to agents in production.
