Loops HouseExploit Escrow

NEAR HACKATHON
Links
Team
1 member- DAOwner
Daniel asaboro
Overview
A researcher with a working exploit can't prove it works without revealing it, and a vendor won't pay for an unproven claim. Exploit Escrow is a NEAR escrow for that exchange, tested on our deliberately vulnerable fixture, acme-notes-api. Testnet only.
What is different is that the exchange's fairness is machine-checked, not argued. The enclave-signed escrow binds the attestation to the exact ciphertext sold (committing AEAD plus a REPORTDATA commitment), so "attest A, sell B" fails. Its atomicity, binding and no-replay properties are checked by TLA+ (with failing negative controls), ProVerif, EasyCrypt (CMT-4 binding reduced to collision resistance) and Kani (6 harnesses). On NEAR testnet, a 10/10 adversarial suite reverts every attack and settles only the valid release (eescrow82901.testnet).
The verdict "this exploit works" is a RISC Zero zkVM receipt, and that receipt is verified on-chain on NEAR (eroc46845v.testnet). escrow-zk (ezkesc61787.testnet) pays only on a valid pass=true receipt bound to the target and the exploit commitment.
On the pause path, the exploit runs in an OutLayer Intel TDX worker. Oracle v2 (eorc257734.testnet) pins the verifier hash and stores each verdict. On a confirmed critical verdict it computes a pause digest on-chain, the guardian (eguard86568.testnet) gets a NEAR chain-signatures MPC signature over that digest, and a Solidity PauseGuard verifies it with ecrecover (forge 9/9).
We published a self-audit (docs/AUDIT.md). In it, we found that the pause path was an open MPC signing oracle: anyone could get the guardian key to sign any hash. We showed this on testnet, fixed it (A1), and the replayed attack now fails. Still open and disclosed:
- A3: escrow-zk pays on proof of possession, not delivery.
- A4: the exploit commitment is unsalted.
- A5: release keys are seller-chosen.
- A8: the separate Groth16 demo tier uses a public seed. A live OutLayer TDX re-run of the fixed verifier is pending an OutLayer testnet outage.