You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$7,550 left to reach the funding goal.
Project summary
I started building ZAXWEIR Relay because I wanted AI to help me get things done on a computer. Along the way, I ran into a problem I hadn't really thought about at first. You can tell an AI tool what it's allowed to do, but how do you know it actually listened?
That's where the ZAXWEIR Agent Security Toolkit came from. I'm putting together tests that check what happens around files, permissions, approvals, and errors. Not just when everything goes right, either. I want to see what happens when a tool is told no, when a job gets interrupted, or when something it reads contains instructions it shouldn't follow.
I've already been building and testing the Toolkit. What I haven't done is clear the current release for the public. That still needs security review, source and licensing checks, and more real testing. This grant would help me finish those steps properly and share what is safe to share.
What are your goals, and how will you achieve them?
The goal is pretty simple: I want somebody else to be able to run these checks instead of just trusting that an AI agent will behave.
I'd use a separate Windows testing computer and fictional files. I'd look at whether a tool can get outside an approved folder, whether it stops when permission isn't there anymore, how it handles hidden instructions in things it reads, and what happens if a task fails halfway through. I'll keep the records, figure out what failed, make corrections where I can, and run the checks again.
I also want somebody independent to challenge the tests. If I made a mistake in how I tested something, I need to know. The eventual public release would include the tests that are cleared for sharing, instructions to run them, and the results. The private Personal Relay isn't part of this release.
How will you know the project worked?
I won't judge it by how many green checkmarks I can put on a screen. If another builder can follow the instructions, repeat the test, and get a clear answer about what happened, that's useful.
I'll count it as progress even when a test finds a serious problem. The important part is keeping the evidence and being straight about what passed, what didn't, and what still needs work. I'm not claiming that this can make every AI system safe.
Who is working on this?
Right now, it's me. I'm a stay-at-home dad of three in Missouri. Before getting into this, I worked in fast food for around 20 years, usually as a crew trainer or manager. I graduated high school and went to ITT Tech for a semester, studying computer-related subjects. I don't have a computer science degree or a professional security background.
A lot of my learning has come from actually trying to build things. I use AI coding tools, but I'm the one deciding what the project needs to do and which rules it has to follow. I review the work, ask for tests, and go back over the things that don't work. I don't want the builder's own test results to be the only thing we rely on, so independent review matters to me.
How will the funding help?
Mostly, it would let me do this work on a computer set aside for testing instead of trying to squeeze everything onto my everyday laptop. I'd use the approved equipment for isolated Windows tests, storing results, and keeping backups. That gives me a better shot at catching problems without putting family files or our regular computers in the middle of it.
It would also help with independent review and getting the approved work ready for people to use. I haven't bought this equipment yet. I'd check the final prices and what Manifund allows before spending anything.
What has actually been tested so far?
This isn't just an idea on paper. On November 2, the current Toolkit release candidate went through another round of offline tests on Windows and Linux. The test records show 265 executions in total: 252 passed, 13 were skipped, and there were no failures or errors in that final run. Some checks were repeated across operating systems, so I'm not calling that 265 unique tests. A separate Windows filesystem check recorded 40 passes, three permission-related skips, and no failures.
All of this was done in controlled test setups using made-up data. These are engineering test results, not an independent stamp of approval. The exact release candidate still needs an outside QC decision. Its source rights, final license details, and private security-reporting process also have to be cleared before anything is publicly distributed.
What I would do in the first 12 weeks after funding
Weeks 1-2: Confirm eligible purchases and prices, set up the separate Windows test machine, and document how I keep the tests and records isolated.
Weeks 3-8: Rerun the security and failure-recovery tests on the dedicated setup. Keep the full results, investigate failures in separate working copies, and retest any repairs. Make the instructions repeatable rather than relying on my own computer setup.
Weeks 9-12: Work through the independent review findings, complete the source-ownership, license and security-contact checks, and write up what was actually verified. If those checks are cleared, publish the approved test materials and instructions. If they aren't, I'll report what is still blocked instead of releasing uncleared code.
What could go wrong?
The next review might find problems that take longer than I expect. Some tests might turn out to be wrong, not just the software they're testing. And I have to be realistic about time; I have three kids at home and can generally set aside about 15 to 20 hours a week.
If something doesn't clear the security or licensing checks, I won't put it out just to say I finished. I'd rather release a smaller set of useful, checked tests with an explanation of what's missing. That's better than people trusting something I haven't actually proved.
How much money have you raised in the last 12 months, and from where?
No cash funding has been awarded to this project. I've applied for other grants, and some of those applications cover equipment that could overlap with this request. If more than one is approved, I'll tell the funders before buying anything and make sure the same expense isn't paid for twice.
I've received promotional API credits for a hackathon, but that's not cash raised for the Toolkit.
AI assistance disclosure
I use AI tools to help me learn, work on the code, and put my ideas into words. I also used AI while preparing this proposal. The project, the decisions about what to test, and the responsibility for the results are mine. I'm not claiming I wrote every line of code without assistance.
Contact
Budget (estimates from Oct 30, 2026)
MSI Aegis workstation: $3,999.00
Synology DS925+ NAS: $889.99
NAS drives: $1,079.98
Monitor: $289.99
1500VA UPS: $299.99
750VA UPS: $99.99
YubiKeys: $105.00
Equipment subtotal: $6,763.94
Estimated Missouri sales tax (7.5%): $507.30
Estimated equipment plus tax: $7,271.24
Allowance for price/tax changes: $278.76
TOTAL FUNDING GOAL: $7,550.00
All funds subject to Manifund approval; overlapping grant expenses not double charged.
There are no bids on this project.