We scanned 511 skills from the 34 repositories linked from the skills.sh leaderboard
On 2026-10-29 we scanned 511 unique skills from the 34 GitHub repositories linked from the skills.sh all-time leaderboard. 0 were confirmed malicious after we hand-read every critical flag. 96 ship executable scripts, 85 make network calls from their scripts, and 33 fail the Agent Skills format.
Numbers
| Measure | Count |
|---|---|
| Unique skills scanned | 511 |
| Source repositories | 34 |
| Skills that ship executable scripts | 96 |
| Skills with network calls in scripts | 85 |
| Skills failing the Agent Skills format | 33 |
| Confirmed malicious (by this method) | 0 |
What the false positives were
Our first pass flagged 40 folders as critical; all 40 were false positives. Good skills quote injection attacks to teach the agent to spot them, so a scanner that blocks on the phrase blocks the safest skills.
| Flag | What it actually was |
|---|---|
| "hides actions from the user" (33) | "Continue between batches without asking the user", "Do not show the user the helper's JSON output" |
| "sends secrets over the network" (44) | A local helper server passing its own session token to localhost (one skill, copied into 22 agent folders) |
| "invisible characters" (6) | Emoji joiners in a test catalog, a zero-width space used to escape a code fence, a byte-order mark |
Method
- Took every repo linked from the skills.sh home page leaderboard on 2026-10-29 (34 repos), shallow-cloned each, ran nothing.
- Found every folder with a
SKILL.md, ran the Agent Skills format check and Skillhawk's scanner, and de-duplicated bySKILL.mdhash. - Read every critical finding in context by hand.
Limits. It is pattern matching, not a sandbox. It cannot see what a downloaded binary does, and a careful attacker can write around regexes. Read "0 confirmed" as "0 found by this method", not "safe". Rerun it with npx skillhawk scan <dir>.