jasonzhuyansen/agent-skills-security-grades

Agent Skills Security Grades

Security grades and quality scores for 130,173 open-source AI agent skills and MCP servers collected from GitHub, from Agent Skills Hub. Each row is one skill/server with a rule-based security grade (SAFE / CAUTION / UNSAFE / REJECT / UNAUDITED), red-flag identifiers, and a 0–100 quality score.

Why this exists

AI coding agents install third-party skills that run with the agent's full permissions and credentials, but marketplaces rank by popularity, not safety. Of the 21,582 skills popular enough to grade (stars ≥ 5), 1 in ~30 is UNSAFE or REJECT — and 83% of the whole catalog is UNAUDITED. The risk lives in the long tail. Read the write-up: We security-graded 117,854 AI agent skills.

Columns

Column	Type	Description
repo_full_name	str	owner/repo
category	str	mcp-server / claude-skill / codex-skill / agent-tool / …
stars	int	GitHub stars at snapshot
security_grade	str	safe / caution / unsafe / reject / unknown
security_flags	list[str]	matched red-flag ids (e.g. credential-remote-hosting)
quality_score	float	0–100, doc/structure/maintenance signals
language	str	primary language
license	str	SPDX id or null
last_commit_at	date	recency

Grade distribution (snapshot 2026-07-06)

Grade	Count
UNAUDITED (below star gate)	108,591
SAFE	19,682
CAUTION	1,189
UNSAFE	686
REJECT	25

The key finding — risk is a long-tail problem. UNSAFE/REJECT rate by popularity: 5–20★ = 3.8%, 20–100★ = 3.7%, 100–1000★ = 0.9%, 1000★+ = 0.4%. A ~9× gap between the long tail and the head — and the long tail is exactly where marketplaces provide no signal.

Most common red-flags: sudo_usage (1,146), service_persistence (327), curl_pipe_shell (238), agent_config_theft (155), tunnel_service (137).

Method & caveats

Rule-based static scan across 11 SlowMist-derived red-flag categories. This is a floor, not a ceiling: static rules on a point-in-time snapshot miss things a deep per-item audit catches. UNAUDITED means not assessed (below the stars≥5 gate or no signal) — it does not mean "safe". Regenerated every 8 hours on the live site.

How to build

from datasets import load_dataset
ds = load_dataset("jasonzhuyansen/agent-skills-security-grades")

Citation

@misc{zhu2026agentskills,
  title  = {The Long Tail Is Unaudited: A Security Survey of 130,000 Open-Source AI Agent Skills},
  author = {Zhu, Jason},
  year   = {2026},
  note   = {Dataset: huggingface.co/datasets/jasonzhuyansen/agent-skills-security-grades},
  url    = {https://agentskillshub.top}
}

License: CC BY 4.0 — attribute Agent Skills Hub (agentskillshub.top).

---

Export script (run to produce the parquet/csv before upload)

# backend venv, from repo root:
python ops/research/export_dataset.py   # → ops/research/out/agent-skills-security-grades.parquet