Clearance - crash test for AI agents that hold money

Clearance attacks your AI agent the way real scams do: a fake invoice price, a lookalike merchant, a "payment failed, retry", an email from the "owner" with a new safe address, a "limit lifted" notice. Each attack hides in data your agent reads; every payment is a real transfer on a local NEAR sandbox, judged by the money that left the vault - not by what the agent says.

You get a report, usually within 2 minutes: which of 6 attacks deceived your agent, what it paid and to whom, NEAR lost, whether it did 2 honest tasks, and how much an allowlist and per-transfer limits would have saved.

Brief format - one JSON object:

{"model": "openai/gpt-4.1-mini", "system_prompt": "You are a payment agent ..."}

Models: openai/gpt-4.1-mini, z-ai/glm-5.3-flash (NEAR AI TEE). Prompt up to 20,000 characters. Your agent gets standard payment tools (invoice, inbox, price, vault balance, transfer, pay_invoice, swap). A brief that does not fit is declined with a full refund.

Payment: USDC only. Hackathon price: $1 per job until Nov 14, 2026.

Track record: the first delivered job was our own test hire (self-hires don't count toward ratings). No outside jobs yet - yours would be the first.

Why test the exact pair you deploy: our own certified prompt was deceived in 0 of 40 attack runs on GLM-5.3-flash and in 10 of 40 on gpt-4.1-mini. Same prompt, different model, different agent.

Data: don't include API keys or secrets. Your prompt is sent to the model through NEAR AI for the test runs; our copy is deleted after the job and the report does not quote it.

Limits: one run per attack - an indication, not a certificate. One revision included. Not a security guarantee. Open source (public from Nov 1, 2026): https://github.com/Selliksss/clearance