Agent Skill Validator
Scan any agent skill repository for prompt injection, malicious scripts, hardcoded secrets, and quality issues. Get a safety grade instantly.
What this tool checks
- Prompt injection and jailbreak patterns
- Hidden instructions and zero-width characters
- Data exfiltration and credential harvesting
- Dangerous shell commands in scripts
- Hardcoded API keys and secrets
- SKILL.md structure and agentskills.io spec compliance
Scan a skill in seconds
Paste a GitHub repo, tree, or raw URL containing a SKILL.md file.
Supports github.com/user/repo, /tree/branch/path, /blob/branch/SKILL.md, and raw.githubusercontent.com URLs
What the Agent Skill Validator does
The Agent Skill Validator checks an AI agent skill before you install it. A skill is a folder with a SKILL.md file (the format used by Claude Code, Codex and other agents that follow the agentskills.io spec), plus optional scripts the agent can run. The validator reads SKILL.md and up to 5 script files (.py, .js, .mjs, .ts or .sh, each under 50 KB) from the same folder of a public GitHub repository.
It runs pattern-based safety checks for prompt injection and jailbreak phrases, hidden instructions and zero-width characters, data exfiltration and credential harvesting, dangerous shell commands, obfuscated code, untrusted installs and hardcoded API keys, plus quality checks against the SKILL.md format. You get a score out of 100, a letter grade and a list of every failed check. The skill is fetched through the GitHub API and is never executed.
How to use it
Copy the GitHub URL of the skill. A repository root, a /tree/branch/path folder link, a /blob/branch/.../SKILL.md file link and a raw.githubusercontent.com link all work.
Paste it into the Skill URL field. Optionally switch on Extended Review for an extra rules-based pass over the instructions.
Click Validate and wait a few seconds while the files are fetched and scanned.
Review the grade and severity counters, then read Safety Findings first and the Quality Checks list after that.
Without an account you can run 1 scan per day. Signed-in users get 5 scans per day. Only public repositories can be scanned.
What the results mean
Score and grade
Every scan starts at 100. Failed checks deduct points by severity, with critical findings costing the most. A+ is 90 or higher, A is 80+, B is 70+, C is 60+, D is 50+, and anything below 50 is an F.
Critical and high safety findings
Treat these as do not install until you have read the flagged lines yourself. Telling the agent to ignore earlier rules, hiding text in invisible characters, sending data to unknown URLs or piping a download into a shell are things a legitimate skill rarely needs.
Quality findings
Missing or invalid YAML frontmatter, a missing or badly formatted name, a missing or overlong description, or a SKILL.md over 100 KB will not make a skill dangerous, but they make it less likely the agent loads and triggers it correctly.
A clean grade is not a guarantee
Pattern checks can miss well-obfuscated payloads, and files in subfolders or beyond the 5 script limit are not scanned. Read the code of any skill that runs commands or touches credentials before trusting it.