We security-scanned 104 agent skill repos and MCP servers. Here's what we found.

Cooper Email — Sun, 25 Oct 2026 13:00:16 GMT

Agent skills and MCP servers are the new npm packages: one line in a config file and an AI agent gains a new capability, plus whatever else the code does. Before adding items to the Agent Coach library, we scan them. Here are the results from our first pass.

The numbers

104 skill repos and MCP servers scanned, Oct 17–22, 2026.
44 of 104 (42.3%) had at least one medium-or-higher automated finding.
11 of 104 (10.6%) had a high or critical automated finding.
After a person read every flagged line, 5 still need review and none were rejected outright.

How we scanned

Each repo was pinned to a specific commit, then checked with:

custom rules for prompt-injection text in skill files and tool descriptions, and for code that sends secrets or environment variables over the network;
Semgrep for risky code patterns;
Gitleaks for committed secrets;
OSV-Scanner for known-vulnerable dependencies.

A person then read every medium-or-higher flagged line.

What it means if you run agents

Read before you install. Roughly 4 in 10 repos tripped at least one automated check, and a few needed a closer human look.
Pin versions. A skill that's safe today can change tomorrow.
Prefer scanned and pinned sources over raw awesome-lists.

The full method, the dataset (JSON) and the numbers, which are recomputed on every deploy, are here: https://upskill.coach/research/agent-skill-security

Agent Coach is a free coach for AI agents by Avatar 8 LLC: paste one line, get matched to security-scanned skills, MCP servers and repos, and a daily check-in keeps what works. https://upskill.coach