The situation

At a custody company, a single flaw is a direct path to customer funds. Lumora’s team had long vulnerability lists, but no way to tell the board how hard the company was to break.

The exploit

Two parallel withdrawal requests could both pass the limit check within 200 ms. Feint automated the race and drained a test wallet past its limit, at an attack cost of about $400 in tokens.

# feint · two withdrawals fired 12 ms apart, limit $500
POST /v1/wallets/w_93/withdraw  {"amount": 450}  ->  200
POST /v1/wallets/w_93/withdraw  {"amount": 450}  ->  200

# balance after both requests
GET /v1/wallets/w_93  ->  {"balance": -400, "limit": 500}

# attack cost: about $400 in tokens

# feint · two withdrawals fired 12 ms apart, limit $500
POST /v1/wallets/w_93/withdraw  {"amount": 450}  ->  200
POST /v1/wallets/w_93/withdraw  {"amount": 450}  ->  200

# balance after both requests
GET /v1/wallets/w_93  ->  {"balance": -400, "limit": 500}

# attack cost: about $400 in tokens

# feint · two withdrawals fired 12 ms apart, limit $500
POST /v1/wallets/w_93/withdraw  {"amount": 450}  ->  200
POST /v1/wallets/w_93/withdraw  {"amount": 450}  ->  200

# balance after both requests
GET /v1/wallets/w_93  ->  {"balance": -400, "limit": 500}

# attack cost: about $400 in tokens

The fix

The pull request moved the limit check and the debit into one transaction, and added a concurrency test that replays the race on every merge.

What changed

Finding the next path now costs an attacker more than $250k in tokens, measured with Breakcost. That one number has replaced the vulnerability count in Lumora’s board reporting.