The five-minute skill check
Before you install an agent skill from someone you don’t know
1
Read the SKILL.md
The file your agent will read as instructions, every line. Not the README of the marketplace.
2
Search for addresses
Does every domain really belong to the vendor? Is the spelling exactly right? Every URL is a place where something different may sit tomorrow.
3
Look into the scripts
Not only the Markdown files. And everything the skill wants to install: “prerequisites”, “setup”, “install it globally”.
4
Origin over popularity
Who submitted it? Stars belong to the repository, not to a single skill in it.
Three rules that matter more than the check
- A skill that fetches instructions from the internet cannot be reviewed once and for all. What gets executed is whatever sits behind the link right now.
- Treat untrusted skills as inspiration. Let your agent rewrite them in its own words, then read the result.
- Keep the safety nets on. Know which mode your agent starts in (Claude Code: auto mode). Sandbox on, no Turbo,
--dangerously-skip-permissions, bypass permissions or--yolowith a new skill, and read every permission prompt. In Claude Code, checkallowed-toolsin the frontmatter.
From the talk “Dangerous AI Skills” by Johannes Hoppe · Slides, prompt and sources: talks.agentic.schule/dangerous-ai-skills/materials/