Skip to content

Is that agent file safe to load?

AI agents follow instruction files - CLAUDE.md, SKILL.md, AGENTS.md - the way computers follow code. Snyk found security flaws in 37% of the public agent skills it checked.

A Trojan horse beside ChatGPT, Claude and agent instruction files

Check a skill before you install it

Everything is read in your browser. Nothing passes through us, nothing is stored.

What it checks

Invisible instructions
Unicode tag and direction-override characters - text an AI reads but you can't see.
Download-and-run commands
Instructions that fetch something from the internet and execute it in one step.
Hidden or encoded commands
Base64 blobs and encoded PowerShell - decoded on the spot and re-checked, so you see what they actually say.
Data sent out
Uploads to webhook collectors, paste sites and other known exfiltration endpoints.
Secrets in reach
Lines that touch stored credentials or keys, flagged harder when the network is on the same line.
Covert behaviour
Phrases like 'don't tell the user' - and instructions that copy themselves into your other agent files.

What it can’t tell you

This is a red-flag check, not an antivirus. It matches the attack patterns documented by security researchers - it cannot catch a novel technique, judge intent, or see malware already running on a machine. “No known red flags” means exactly that, and nothing more. The wider story - fake AI installers, stolen browser sessions, poisoned skills - is in our article on malware dressed as AI tools.

Want the whole environment checked, not just the files?

Where your AI tools came from, what extensions your team installed, whether the agent setup is clean. That audit is the first step of our AI implementation work.

Book a Consultation