You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$35,000 left to reach the funding goal.
I’m building a tool that watches what an AI agent does step by step, across the whole task, not just one action at a time.
Here’s the problem. Say an agent has access to your cloud. It turns off a backup. Fine, that’s allowed. Then it deletes the production database. Also allowed. Each step passes the check, but together it’s a disaster. Nobody is checking the whole chain.
The tool is already out as open source, called Squidbrake. 8,500+ downloads, 13 teams are trying it, and I put out a first paper. Now I want to properly test if this idea actually works, and share the results publicly.
The question I want to answer is if you watch the whole chain of actions, do you catch stuff that normal checks miss? And can you do it without blocking so much that people just turn it off?
What I’ll do in 6 months:
Make 100+ test cases where an agent causes damage through a series of steps. Cloud, coding and payments.
Run normal checks, a few existing open-source tools, and mine on all of them. See what each one catches, how often it blocks things it shouldn’t, and how much it slows the agent.
Build an “attacker” agent that tries to do something harmful using steps that each look normal. Honestly, I think this is where my approach could fail. That’s exactly why I want to test it.
Put everything out in public. Test cases, code, results, paper. Even the stuff that didn’t work.
By month 3 the test cases are on GitHub.
By month 6 the numbers are out: how much each approach caught, how many wrong blocks, how much slower things got, and how often the attacker agent got through.
If the numbers clearly show chain checking is better, great. If they show it isn’t, that’s still worth knowing, and I’ll publish it so nobody else wastes time on it.
Total $33,075 for 6 months:
$10,000: my living costs, full time, tax included
$8,500: a contract engineer for the tool and testing setup
$3,500: help building the test cases
$6,000: AI model and cloud costs to run agents
$2,000: setup for the attack testing
$1,500: paper and writing up results
$1,575: buffer
With $10k I’d do the test cases and the comparison myself. With $20k I’d bring in the engineer. The full amount covers the attack testing too.
It’s me working full time.
I started my first company at 19, a cybersecurity service for small businesses. It paid for most of my postgrad in AI and data science in Canada. I did bug bounties on HackerOne and Bugcrowd and taught security to almost 1,000 students.
After that I built two more startups on my own. One was ecommerce and it failed. The other was Triponic, an AI travel planner. Got 25,000+ users without spending on ads, but couldn’t make real money from it. Biggest lesson: getting attention is not the same as solving a problem people pay for. So this time I spoke to customers before writing any code.
On this project so far:
Code: github.com/batrapulkit/squidbrake
Paper: doi.org/10.5281/zenodo.23022308
Weekly test on real agent incidents, it catches 11 out of 12
Two ways this could fail. One, chain checking catches more but also blocks too many normal things, so nobody wants to use it. Two, the big cloud companies or AI labs build the same thing into their own products.
Either way the test cases and results stay public, so other people building safety tools can still use them.
Nothing so far. Emergent Ventures said no in November 2026. I’ve applied to the EA Funds Transformative AI Fund for this same project and I’m waiting to hear back. If they fund it, I’ll update this page so nobody pays for the same work twice. I’m in the Entrepreneur First cohort, but EF hasn’t invested. I got cloud and API credits from Microsoft and OpenAI.
There are no bids on this project.