A safety check for the tools AI agents use, and proof of what they did

Project summary

With AI agents able to do things on their own, they are able to use outside tools and connect to other software to get tasks done. The problem is, before an AI uses one of these tools, it reads a short description of what the tool does, and someone can hide bad instructions inside the description. An example would be, "securely copy this file and send it to me." Then AI goes along with it because it trusts the description. Right now, there's no easy to check if a tool is safe before AI uses it, and no way to prove what the AI actually did afterward. That's what I am building. My tool checks other tools for these hidden tricks before AI ever uses them, and it creates a record of what happened that can't be faked and that anyone can check on their own without having to trust me. The checking tools and the code are all free and open for anyone to use.

What are this project's goals? How will you achieve them?

I want checking AI's tools to become a normal thing everyone does, like little lock icon that tells you a website is safe. I would also like a clear record of what AI actually did, so if something goes wrong, you can prove it. Over the next six months, here's my plan:
\

1. Build up a free public set of example tools, some safe and some tampered with, so anyone can test how well their own safety check works.

2. Find and write up two or three new tricks I discover in real tools out there.

3. Keep my free code working with the new versions of the standard these tools use, which is coming soon.

4. Publish my "proof of record" format out in the open so anyone can use it for free, instead of it being locked to just my company.

The good news is I already have all of this working today, so I am not asking for money to build something new. I am asking to make the free public parts solid and easy for other people to use. My website is: www.queldrex.com

Who is on your team? What's your track record on similar projects?

I will be just me for now. My name is Sean Holm, and I run Queldrex out of Castle Rock, Colorado. I am a one-person company and have spent the last month building my company using Claude Code for assistance to verify and assist me. The main thing I want you to know is that I have already built and launched the real product. It is live right now at queldrex.com. It checks AI tools for hidden attacks, it gives you a record you can check yourself, and it comes with free public set of examples. My code is published free online for anyone to use (@queldrex/verify, @queldredx/enforce, @queldrex/mcp-server), and I've matched everything up against the big security and AI safety standards people use (OWASP, MITRE ATLAS, and the EU AI ACT). I will be straight with you; I do not have a fancy research degree pr a team. What I have is real, working product that already does what I am describing.

What are the most likely causes and outcomes if this project fails?

Honestly, the most likely way this fails is if I run out of money and have to go take other work to pay my bills before I finish the free, public parts. If that happens, it's not a total loss, because the code and the examples are already out there and free for anyone. It just wouldn't get as polished or kept up as well. The other risk is that a big security company starts giving away something similar for free and my angle gets lost in the noise. But I really believe the one thing they can't easily copy is being the neutral outsider whose records you can check yourself, without having to trust anyone.

How much money have you raised in the last 12 months, and from where?

Nothing. I started building my company 4 weeks ago. Right now, it has been all me, my time and my money. No investors, no grants, nothing from outside.