You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$25,000 left to reach the funding goal.
Clearance is a crash test for AI agents that hold money. It hides the scams that work on people (a fake payee, a padded invoice, a fake "payment failed, try again") inside the data an agent reads, lets the agent pay with real transfers on a NEAR blockchain sandbox, and counts how much money it lost.
Agents are getting wallets now, and nobody checks if they can be scammed before they get the money. I want a crash test to be a normal step before an agent gets a wallet, the way cars get crash tested before people drive them.
How I plan to get there:
keep the test open and free to run. Anyone can paste their agent's prompt at agentcrashtest.com/test, or try to beat our certified agent at agentcrashtest.com/challenge
run real agents for their owners and publish the results when they agree
keep a public table of which models fall for which lies, updated when new models come out. Right now GPT-6 and Claude Opus 5.5 both paid twice after one fake "payment failed" message
give each agent that passes a certificate on chain that other contracts can check, and revoke it when someone finds a lie it falls for
teams send their own agents to the test (goal: 50 in 6 months)
some of them pass and get certified (goal: 10)
wallets, marketplaces or audit firms use the results or the certificate
people beat the certified agent in the challenge and we publish how
For 6 months:
server for the runs: $1,000
running frontier models (GPT, Claude, Gemini) for the public table: $5,000
marketing, so people building agents actually find it: $2,000
my time, so I can work on this full time: $17,000
With the minimum ($5,000) I keep the server and the model runs going and keep the table public.
Just me. I'm 18 and I build with AI coding tools. I built Clearance alone in one week during the NEAR hackathon: the contract, the test runner and judge, a CI signer in GitHub Actions, a Telegram bot and the site. Everything is in the open repo, with 1847 runs anyone can replay: github.com/Selliksss/clearance. One startup (UPRAIL) shipped a fix after my report. Before this I built a token tracker and a tool that reuploads YouTube videos to a Russian video site, which made $40,000 in half a year.
Most likely: agent owners don't want to test their agents, or don't want the results public. Or new models get much better at this and the test gets old. Even then the attacks, the runs and the code stay open, so anyone can reuse them or build on them.
None. Everything so far is paid from my own money.