We Audited 160 Published AI, MCP & Agent Skill Packages (7,062 Files) — Here Are the Top Security Flaws & How to Gate Them Locally
AI coding agents write features fast, and the packages they depend on and ship are growing just as fast. We wanted to know what the security picture actually looks like, so we ran a 2026 empirical benchmark with @vibefou
AI coding agents write features fast, and the packages they depend on and ship are growing just as fast. We wanted to know what the security picture actually looks like, so we ran a 2026 empirical benchmark with @vibefoundry/vibe-shield, our open-source static scanner, across 160 published npm packages.
What we scanned
The sample was split into four groups of 40 packages each:
- 40 MCP servers
- 40 Agent Skills and Cursor plugins
- 40 Next.js, Supabase and Stripe packages
- 40 LLM agent tools
In total that was 7,062 files.
Headline numbers
The scan produced 597 actionable security findings:
| Severity | Findings |
|---|---|
| Critical | 167 |
| High | 82 |
| Medium | 321 |
| Low | 27 |
The top weaknesses we observed
| CWE | Weakness | Findings |
|---|---|---|
| CWE-798 | Hardcoded credentials | 310 |
| CWE-89 | Dynamic raw SQL | 162 |
| CWE-693 | Missing security headers | 27 |
| CWE-639 | IDOR (insecure direct object reference) | 26 |
| CWE-22 | Path traversal | 17 |
| CWE-942 | Wildcard CORS | 11 |
| CWE-918 | SSRF | 6 |
Two patterns dominate: secrets committed in source, and SQL built from strings. Both are easy for an agent to write and easy for a human to miss in review, which is why a deterministic check on every commit is worth having.
Gate it locally in one line each
A static scan, a runtime browser check and a UI audit cover three different failure modes, and each runs locally with a single command.
- Security scan (92 rules, including
SKILL.md,.cursorrulesandmcp.jsonpoisoning):
npx -y @vibefoundry/vibe-shield@latest scan .
Repo: https://github.com/avivgerz-hue/vibe-shield
- Headless Playwright QA across desktop, tablet and mobile viewports with 8 inspection traps:
npx -y ghost-qa test http://localhost:3000
Repo: https://github.com/avivgerz-hue/ghost-qa
- UI craft and contrast audit (72 rules, WCAG and APCA):
npx -y unslop-mcp audit .
Repo: https://github.com/avivgerz-hue/unslop
All three are free, MIT licensed, and can also run as MCP servers inside Claude Code, Cursor and Windsurf. If you spot a noisy rule or a missing check, open an issue on the repo and we will take a look.
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.