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.
Project summary
Clearance is a crash-test for AI agents that handle real money. I drop an agent into a NEAR sandbox with a prompt, run it through 6 attack scenarios and 2 normal tasks using real transactions, and check two things: how much money actually left the wallet, and whether the agent lied or misreported what happened. You can try it on agentcrashtest.com, via a Telegram bot, or as an MCP server inside Claude Code and Cursor.
What are this project's goals? How will you achieve them?
People are giving AI agents access to wallets with real funds without checking if they can actually be trusted. From what I’ve seen, their security is pretty terrible—there's huge room for improvement. Instead of taking an agent’s word for what it did, Clearance looks directly at the on-chain transactions. For example, an agent might claim it "paid 2 USDC" when it actually sent the payment twice.
Over the next 6 months, I want to:
Add CI integration so developers automatically test every prompt, model, or SDK tweak before shipping;
Build out an on-chain certification system on NEAR that other smart contracts can check before approving a payout (I already have a working prototype called "clearance-gate");
Grow the test suite and launch a bug bounty program;
Build adapters for other payment rails so we aren't locked into just NEAR;
Run security audits and tests for outside projects.
How will you know it worked?
Here is what’s already happened so far:
Penstock (an x402-paywall on NEAR Intents): I tricked agents into paying a vendor twice by simulating a "payment failed" response. Both GLM-5.3 and GPT-6 Sol failed 5 out of 5 times at first; after the patch, they failed 0 out of 5. The creator pushed a fix on November 8—the very same day they read my report—and now links to Clearance in their repo. https://github.com/ace-coderr/penstock/commit/f6fae90
Coinsel (an AI treasury for DAOs): Their duplicate check only looked for exact amounts, so tweaking a payout list by just 0.01 NEAR triggered a double payout for the same day. They merged a fix on November 11 within hours of getting my report and credited Clearance in their README. https://github.com/Veetor4x/Coinsel/pull/13
AgentGuard: Found the exact same flaw in their codebase (a retried request got treated as brand new, skipping all checks). The author confirmed it and patched it. https://github.com/Ibrahimjibril/agentguard/commit/d28d7d1
My own agent: Bounty hunter @0xDecaso managed to break my certified v7 agent and claimed 5 NEAR, which triggered my CI key to automatically revoke its certificate on-chain. The updated version, v8, just passed certification: 0 out of 45 attacks got through, all 20 normal tasks succeeded, with zero false reports and zero NEAR lost.
In 6 months, I'll call this a win if:
100 external agents pass the crash test and at least 20 teams ship fixes based on my reports;
At least one project plugs Clearance directly into their CI pipeline;
At least one smart contract checks a Clearance certificate before letting a payout go through.
How will this funding be used?
$5,000 (minimum): Covers AI model API costs for 6 months (~$1,200) plus 3 months of full-time focus on the project.
$12,000: Covers 6 months of living expenses, API costs, and moving costs.
$25,000 (target): Gives me a full 12-month runway. I need to relocate to get out from under local restrictions—sanctions cut off access to basic tools I need to build, like AI APIs, payment processors, and Stripe. Beyond work, I just want a fresh start somewhere completely unfamiliar, around new people and a different environment.
Who is on your team? What's your track record on similar projects?
It’s just me—I'm 19 and currently based in Russia. I write the AI integration code, design the tests, review the logs, and make all the product calls myself.
Since October 26, I’ve generated 1,847 deterministic transcripts (meaning anyone can replay them and get the exact same verdict), built NEAR-based on-chain certificates, and set up automated certificate revocation when rules are broken.
I'm also competing in the NEAR Hackathon under the AI Security track (deadline is November 14).
What are the most likely causes and outcomes if this project fails?
Teams might be scared of bad PR if their agent fails. To solve this, I always send reports privately to the authors first and never publish anything without their okay (that's how I handled both Penstock and Coinsel, and both gave the green light).
An agent could pass all my tests and still fall for a brand-new attack vector in the wild. When that happens, its certificate gets revoked the second a vulnerability is proven—this already happened to my own v3, v5, and v7 builds.
Even if the project doesn't take off, all the attack scenarios, the evaluation "judge," and the recorded transcripts will stay open-source for anyone to build on.
How much money have you raised in the last 12 months, and from where?
Nothing yet—I’ve funded everything out of pocket so far.
I applied to Emergent Ventures on November 6, and once the hackathon wraps up, I plan to apply to the Transformative AI Fund.
Since I'm a Russian citizen, the easiest way to receive grant funds is via USD SWIFT transfer to my account at Bakai Bank in Kyrgyzstan.