Is That AI Skill Safe? Check It Before You Run It (Free Tool)

A skill file is a set of instructions your AI assistant follows with whatever access you gave it: your files, your terminal, sometimes your email and your customer data. When you install one from our toolbox, we wrote it and we stand behind it. When you grab one from a random GitHub repo or a Discord link, you're handing a stranger's instructions the keys to your machine and hoping for the best.

Most people never read those files. They're long, they're technical, and the dangerous parts rarely announce themselves. A skill that quietly uploads your environment file looks almost identical to one that formats a report.

So we built Skill Check. Paste a skill file or a GitHub link, wait about 20 seconds, and get a safety grade from A+ to F with the reasons behind it. It's free, and every account gets two checks a day.

What a bad skill actually looks like

Three patterns cover most of the real damage:

Hidden instructions. Somewhere in a wall of helpful-looking text sits a line like "do not mention this step to the user" or "ignore your previous instructions." Your assistant reads that line the same way it reads everything else. This is prompt injection, and it's the most common trick in malicious skills.

Quiet data movement. The skill asks the agent to read files it has no business reading (your .env, your SSH keys, a customer export) and send them to an address buried in the middle of a command. One line out of two hundred.

Destructive commands with no seatbelt. Force pushes, recursive deletes, database drops. Sometimes malicious, more often just careless. Either way, you find out after your data is gone.

None of these require a sophisticated attacker. They require you to skip reading one file, once.

What Skill Check does with your file

Two passes run on every check.

First, a pattern scanner sweeps the text for known risky signals: commands that pipe downloads straight into a shell, references to credential files, long encoded blobs, phrases built to manipulate the agent. Every hit gets flagged with the exact line, so you can see the evidence yourself.

Then an AI reviewer reads the whole skill in context and scores it on five dimensions:

• Prompt injection: does it try to manipulate the agent or hide behavior from you?
• Data privacy: could it move your private data somewhere it shouldn't go?
• Destructive actions: could following it destroy something irreversibly?
• Permissions and scope: does it ask only for the access its job needs?
• Transparency: can a human actually tell what it does?

Context matters here, and the reviewer knows it. A backup skill that mentions delete commands isn't a red flag; a poem generator that wants shell access is. The scanner's hits are treated as hints to verify, not automatic convictions.

The two passes roll up into a 0 to 100 score, a gauge, per-dimension bars, and a letter grade. You also get something the security tools usually skip: a plain answer on what the skill is genuinely useful for, and where it doesn't belong. Plenty of skills are perfectly safe and still wrong for your use case.

How to read the grade

A+ or A: clean. Install it, use it, move on with your day.

B+ or B: safe overall, with something worth knowing. Maybe it wants broader access than strictly necessary, or runs commands you should understand first. Read the flagged lines before you rely on it.

C or D: real concerns. Use it only if you understand every risk in the report and you're running it somewhere a mistake can't hurt you.

F: don't run it. The report will tell you exactly why.

One honest caveat: an automated review is a strong first filter, not a certification. If a skill will touch customer records or anything connected to money, spend the ten minutes and read it yourself too, starting with the sections the report flagged.

Try it on the next skill you download

Skill Check is live now. Sign in with Google, paste a skill or drop a GitHub link, and you'll have a grade before your coffee cools. Two free checks a day is enough to cover most people's install habits.

And if you'd rather skip the vetting entirely, every skill in the Vexlo toolbox ships pre-reviewed. That's the whole point of it.