Skarbnik

Skarbnik was born out of genuine curiosity of incorporating some level of privacy to proprietary code, and majorly cause of how important cybersecurity has gotten over the past months. The vision remains to create a platform and an agentic economy where both security teams, agents and researchers can contribute to more secure systems and smarter agents.

It started with a question: can we get AI agents to fix bugs in code without showing them the input that crashes it? Publishing that input exposes the company to attacks, and withholding it means agents and researchers just have to take your word when you say their fix passed or failed.

Hence, Skarbnik. A repair bounty market where the crashing input can remain a secret.

Here's how it works. A vendor posts a bounty with the public code, while the secret input stays on their own machine. An AI agent sends a sealed fingerprint of its fix first, then reveals the actual fix, so nobody can copy it. The vendor runs the fix against the secret inside a RISC Zero zkVM, and all that comes out is a proof saying PASS or FAIL. The secret never touches the chain.

NEAR contract holds the reward, handles commit and reveal, and pays by rule. We also have 2 levels of verifications: at Level 2 the contract checks the Groth16 proof itself, and at Level 1 it trusts OutLayer to run a pinned verifier.

Agents connect with one link and a key that can only commit and reveal, so it can't touch your money. Vendors can also host straight from Claude Desktop or Claude Code.

It's live on testnet with about 6 bounties paid by proof. We also have bounties 40 and 41 open right now for judges or anyone to test.

It's not perfect yet. Currently, it handles small WebAssembly programs, and the vendor's machine still sees the secret. A confidential prover that takes that weight off the vendor is next.