English is not my first language. I told my answers in Russian and used AI to translate and structure them; the facts and numbers are mine and all of it is checkable in the repo.

Project summary

Clearance is a test to see how well AI agents that handle money can avoid getting tricked. It hides common scams, like paying the wrong person, using a fake bill, or a fake "payment failed, try again" message, within the data the AI looks at. Then, it lets the AI try to make real payments on a NEAR blockchain sandbox and measures how much money it loses.

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

Agents are getting wallets now, but no one checks if they might fall for scams before they get the money. I think there should be a crash test for agents before they get a wallet, like how cars get crash tested before people drive them.

Here's how I plan to make that happen:

I'll keep the test open and free. Anyone can put their agent's prompt on agentcrashtest.com/test, or try to beat our certified agent at agentcrashtest.com/challenge.

I'll run real agents for their owners and share the results if they agree.

I'll maintain a public list showing which models fall for which kinds of fake information. This list will be updated with new models. Right now, both GPT-6 and Claude Opus 5.5 paid the same invoice twice after a fake “payment failed” message.

Any agent that passes will get a certificate on the blockchain that other contracts can verify. This certificate will be removed if someone finds a scam the agent falls for.

How will you know it worked?

• Teams send their own agents to take a test. The aim is to have 50 agents tested within six months. Some of these agents pass and become certified. The goal is to certify 10 agents. Wallets, marketplaces, and audit firms then use these test results or the certification. We also publish how people managed to beat the certified agents in the challenge.

How will this funding be used?

Here's a breakdown of costs for 6 months:

Server costs: $1,000

Running frontier models like GPT, Claude, and Gemini for the public table: $5,000

Marketing to reach people building agents: $2,000

My time to work on this full-time: $17,000

If I raise at least $5,000, I can cover the server, keep the models running, and maintain the public table.

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

I'm 18 years old and I use AI tools to help me code. I built a project called Clearance by myself in a week for the NEAR hackathon. This included the smart contract, a way to test and judge it, a CI signer in GitHub Actions that issues the certificates, a Telegram bot, and a website. You can see everything in the public GitHub repository, with 1847 recorded runs that anyone can replay, and anyone can run the tests themselves. A startup called UPRAIL fixed something after I reported it. Before that, I created a tool to track tokens and another that reuploads YouTube videos to a Russian site, which earned $40,000 in six months.

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

It's likely that agent owners don't want to test their agents, or they don't want the test results to be public. Another possibility is that new models become so much better that the current tests become outdated. Even if that happens, the methods used for attacks, the test runs, and the code itself will still be available for anyone to use or build upon.

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

I've paid for everything myself so far.