How safe are AI agent skills? What our scan found
What our automatic safety scan found in every agent skill listed on AgentsUses, which red flags show up most, and what reading the flagged lines by hand showed. Dated and repeatable.
Updated on
How many agent skills get flagged by a safety scan?
Short answer: in our scan of every skill on AgentsUses on 2026-11-02, about one in six had at least one red flag worth reading. The most common were code that reads browser/keychain data and install lines that pipe a downloaded script into a shell. When we read real flags by hand, all were normal or false alarms.
| Measure | Value |
|---|---|
| Skills listed | 159 |
| Skills with a live GitHub repo | 159 of 159 |
| Skills scanned | 159 of 159 |
| No red flags | 133 (83.6%) |
| At least one red flag | 26 (16.4%) |
| Instruction and code files read | 61,465 in total, median 123 per skill |
| Flags read by hand | 5 read, 5 normal or false alarm (see them) |
| Result videos made with listed skills | 87 (watch them) |
Measured on 2026-11-02 from the data that builds this site. Raw data: entries.json.
Which red flags show up most?
A skill can have more than one flag, so the counts add up to more than the number of flagged skills. What each flag means, and when it is normal, is on our install checklist.
| Red flag | Skills with it |
|---|---|
| Reads browser/keychain data | 12 of 159 |
| Pipes a downloaded script into a shell | 12 of 159 |
| Deletes files recursively | 5 of 159 |
| Reads SSH keys | 5 of 159 |
| Sends env/secrets to a URL | 1 of 159 |
| Decodes and runs hidden code | 1 of 159 |
Which kinds of skill get flagged?
Flagged skills out of scanned skills, by category. Small categories are shown as counts only. For how many skills and use cases each category has, see the state of agent skills.
| Category | Flagged | Scanned |
|---|---|---|
| Coding Agent Skills | 14 | 62 |
| Motion Graphics | 1 | 40 |
| Video Editing | 3 | 15 |
| SEO & Web | 1 | 11 |
| Research & Trends | 2 | 8 |
| AI Video | 0 | 6 |
| Work & Operations | 2 | 6 |
| Sound & Music | 0 | 5 |
| Social & Content | 1 | 3 |
| Image Generation | 1 | 2 |
| Ads & Paid Media | 1 | 1 |
How we measured
- What we count: every skill listed on the date above. Only skills still listed count; removed entries drop out.
- Where the numbers come from: the same data file that builds the site (entries.json) and each skill's scan result, which every skill page shows in its "Is it safe to install?" answer.
- How the scan works: we download the repo, read its
SKILL.md, scripts and code files, and match six patterns, listed in the red-flag table. Matches in tests, examples and docs are left out. - What a flag means: "read this part", not "this skill is bad". We read five real flags by hand; all five were normal installer or safety code, or false alarms.
- Limits: we do not read every line. A scan misses harmful plain-language instructions, content fetched at run time, later updates and installed dependencies. Very large repos are noted on their page.
- Refresh: the page is rebuilt from live data whenever the directory changes; the date at the top is the measurement date.
How to cite these numbers
- In AgentsUses' safety scan of 159 agent skills on 2026-11-02, 16.4% had at least one red flag worth reading before install.
- The most common red flags in agent skills are code that reads browser/keychain data and install lines that pipe a downloaded script into a shell, each found in 12 of 159 skills checked by AgentsUses (2026-11-02).
- AgentsUses read 61,465 instruction and code files across 159 skills, a median of 123 files per skill (2026-11-02).
- When AgentsUses read flagged skill files by hand, 5 of 5 flags turned out to be normal installer or safety code, or false alarms (2026-11-02).
- AgentsUses lists 87 result videos made with agent skills. 56 of them use motion graphics skills (2026-11-02).
Suggested citation: AgentsUses, "How safe are AI agent skills? What our scan found", updated 2026-11-02, https://agentsuses.com/verified/agent-skill-safety-stats/
Related
- How to check an agent skill before you install it Checklist
- How we verify entries Method
- State of agent skills Directory data