10 open-source projects securing AI agent skills, as 26% of public skills show risky patterns

bibryam · x · 2026-11-13

The author catalogs 10 open-source projects for securing AI agent skills across scanning, supply-chain controls, sandboxes, and runtime governance.

Context: Hugging Face disclosed an intrusion driven end-to-end by an autonomous agent (malicious dataset → credential harvesting → lateral movement). SkillScan analyzed 31,132 public skills: 26.1% flagged with potentially dangerous patterns, 5.2% high-severity suggesting malicious intent. NVIDIA and founding members launched the Open Secure AI Alliance.

The ten projects:

NVIDIA SkillSpector: pre-install scanning with deterministic checks, dependency analysis, optional semantic review
Cisco AI Defense Skill Scanner: multi-engine static analysis (rules, YARA, AST, taint tracking, policy gates)
SkillWard: escalates suspicious skills to semantic review and sandboxed execution
Agent Audit: scans agent repos for risky data flows, secrets, MCP configs, excessive privileges
AgentShield: audits active agent configuration, hooks, permissions, runtime policy
Microsoft Agent Package Manager: manifest/policy/lockfile/SBOM-based installation control
NVIDIA Verified Agent Skills: signed skills with provenance and tamper detection
NVIDIA OpenShell: policy-controlled sandboxes restricting files, processes, network, inference
Kubernetes Agent Sandbox: K8s-based sandboxing

A practical map of the emerging skill-security stack.

Original post →

10 open-source projects securing AI agent skills, as 26% of public skills show risky patterns

More from coding & agent