Skillz Shield
Security screening for Agent Skill packages, with findings and coverage from two scanning engines.
Active development; static security screening. Operation unverified.
Status and source
Availability: Published project. Maturity: Active development; static security screening. Evidence: Documented package screening with Cisco Skill Scanner and NVIDIA SkillSpector; scan results are not safety certification. Delivery: unknown. Source record (reviewed 2026-09-20).
About the project
Find what doesn’t hold.
Security evidence for Agent Skills. Skillz Shield brings Cisco Skill Scanner and NVIDIA SkillSpector into your repository to inspect committed skill packages and give maintainers findings they can review.
Good intentions need a closer inspection.
A skill packages instructions, scripts, and supporting files that an agent may rely on. A useful description alone cannot tell you what those files could expose, change, or request.
Skillz Forge makes reusable capabilities discoverable. Skillz Shield examines their security risks. Its guiding phrase, Find what doesn’t hold.
, comes from The MurderBird: What the Water Kept.
From committed package to reviewable evidence
Select
A committed change selects its skill package and supporting files. Draft packages are included. Application-only changes skip scanning.
Resolve
Official stable scanner releases become immutable, verified artifacts for the run. The Shield integration stays pinned to a reviewed commit.
Inspect
Cisco Skill Scanner and NVIDIA SkillSpector perform static analysis without executing submitted skill instructions, scripts, installers, or dependencies.
Review
Maintainers receive sanitized findings, coverage, source revisions, package digests, and scanner identities. Missing analysis remains visible.
A decision with an evidence trail
Pass
Both engines completed the requested analysis with no high or critical findings. Lower-severity findings may still be present.
Review required
Both engines completed, and at least one high or critical finding needs a maintainer’s review.
Incomplete
At least one package could not be fully analyzed. Findings already obtained remain available; missing coverage is not a passing result.
No result certifies that a skill is safe, useful, mature, or ready for production. A failed scanner resolution or installation can stop before a report exists. Whether a failing scan blocks a merge depends on the consuming repository’s settings.
Focused on skill packages
- Scanning runs in GitHub Actions. The browser displays public evidence and never receives scanner credentials or executes submitted skills.
- Both engines use static analysis without model credentials. NVIDIA’s dependency analysis can contact OSV for vulnerability information.
- Reports omit raw skill snippets and scanner diagnostics. Security screening does not evaluate skill quality, usefulness, or maturity.
- Cisco and NVIDIA maintain their own tools and licenses. Shield integrates their analysis; no vendor endorsement is implied.
Add inspection to your skill repository
The reusable GitHub Action works independently of the websites. Follow the integration guide to pin a reviewed Shield commit, use a disposable GitHub-hosted runner with read-only permissions, and keep scan state outside the source checkout.
The integration guide and repository README maintain the installation steps and update behavior. Review the evidence and coverage for your own packages before relying on a result.