Credit Without Debit
Formal Verification of Banking Payment Systems
University of Massachusetts Amherst
ACM CCS 2026The Hague, Netherlands
Abstract ACM CCS 2026
Most people assume that money credited by a bank transfer has already left another account. Does that hold throughout the transfer? We found a new class of attacks, Credit-Without-Debit: an attacker gains spendable credit at one institution while making the matching debit fail at another. ACH processing separates authorization, internal posting, and interbank settlement, allowing account balances at different institutions to diverge while a transfer is still underway.
We built a formal model of consumer ACH payments from the Nacha rules and used model checking to find security violations. We confirmed three attack instances in live tests at two of the four largest U.S. ACH banks, three card issuers, and three payment providers. The results describe those workflows, not every bank. We disclosed our findings and proposed remedies, and follow-up tests confirm that none of these instances still works. We describe the gap between local payment decisions and transfer-wide guarantees, and how to close it.
01Motivation
Why ACH Payments?#
When a bank transfer looks complete, you reasonably expect the money to stay put. Banks and customers expect every credit to have a matching debit, and pending money to remain unspendable. ACH carries payroll, bill pay, credit-card payments, and bank transfers behind Venmo, PayPal, and Cash App. With 30.7 billion ACH payments at the top 50 U.S. institutions in 2025, these expectations matter in everyday banking.
Each transfer crosses the sender's bank (ODFI), the clearing operator, and the receiver's bank (RDFI) over several banking days. The receiver's bank checks the account only when the entry arrives. The banks make decisions at different times, so everything depends on one question: has the other bank accepted the debit by the time the credit can be spent?
The short version: the Nacha Operating Rules let institutions check authorization once and make provisional credit spendable before the other bank has decided whether to accept the debit. We found that these choices can leave spendable money at one bank while the matching debit fails at another.
02Findings
What We Found#
In the workflows we tested, a bank or payment app released money or restored spending power while the debit meant to fund it was still unsettled. The other bank later refused that debit. The gap let an attacker withdraw provisional funds, keep verification deposits, or spend against a restored card limit without a successful matching debit.
Banking days after the first step. Positions are approximate. Ranges come from the paper.
Gap = credit released minus debit collected. Above $0, that is credit without debit. Dollar amounts are illustrative. Each step follows a trace from the paper.
How the gap becomes a loss
- A customer starts a payment. A bank or app accepts a request involving an account that blocks debits. Account details alone do not establish debit permission.
- An early check looks sufficient. A valid routing number, a successful deposit, or an internal payment posting lets the process continue before anyone knows whether the funding debit will go through.
- Credit becomes available too soon. The customer receives spendable funds, small verification deposits, or a restored card limit while recovery remains uncertain.
- The customer uses the gap. They withdraw funds, spend restored credit, or restart verification to collect more deposits before the institution can recover them.
- The funding debit fails. The receiving bank enforces its restriction. Reversing a ledger entry does not necessarily recover money already spent or retained.
We call this Credit-Without-Debit: one institution releases credit while the matching debit fails at another. We reproduced three instances with ordinary customer actions on accounts we controlled. It does not mean every bank or payment app behaves this way. The common failure was treating an early success, such as a deposit or an internal posting, as proof that the funding debit would settle. Credit should stay unspendable until that debit has passed the other bank's checks.
03Method
How We Found It#
A few successful transfers tell us that those transfers worked. What about every other order of events? Model checking uses a program to try every order of events a model allows, including a debit block that arrives between payment steps. For each security property, it either proves the property holds throughout the model or returns the exact sequence that breaks it. We used nuXmv to search a bounded instance of consumer ACH payments this way, and it returned a breaking sequence for each of the three properties below.
The Nacha Operating Rules are the rulebook U.S. banks follow for ACH payments. The rulebook spans 728 pages. It defines 25 payment categories, called SEC codes, and 87 return codes. It leaves timing choices to institutions, including what counts as acting promptly. We focused on domestic consumer WEB and PPD entries: online payments and prearranged payments such as recurring bills. We modeled the participants and their messages under the relevant rules, then wrote three security properties describing what should stay true throughout a payment.
- Φ₁Continuous authorizationPermission to debit must remain valid from submission through settlement.
- Φ₂Zero-sumCredits and debits must balance separately for micro-deposits and regular payments once a cycle finishes.
- Φ₃Settlement finalityFunds or credit limits become available only after the payment's return window closes.
04Evaluation
What We Tested#
We tested on accounts we controlled. Our sample covered two of the four largest U.S. ACH banks, three card issuers, and three payment platforms. We observed customer dashboards rather than internal bank logs, recording balances, payment statuses, timestamps, and return codes. In every workflow we tested, available credit did not mean the funding debit had succeeded.
For creditBLEED, Bank A made the funds available three days after we submitted the debit, and Bank B later returned it with R16 without debiting the account. All three payment platforms sent micro-deposits to our debit-blocked account, and every attempt to reclaim them failed. All three card issuers restored the full credit limit within two business days, while the funding debit was still unsettled. The debit failed after about six business days. We disclosed the findings on April 16, 2025. creditBLEED and creditRESET no longer worked by May 2025, and the last provider resolved creditMINT by January 2026.
View as table
| Attack | Spendable from | Return or failed debit | Exposure | Invariant |
|---|---|---|---|---|
| A1 creditBLEED | Day 3 in our test (provisional credit) | R16 return, about 2 days later | ≈ 2 days | Φ₁ |
| A2 creditMINT | Day 1–3 (micro-deposits post) | Reclaim debits fail (R08 / R16), Day 4–5 | Never recovered, repeatable | Φ₂ |
| A3 creditRESET | ≤ Day 2 (credit limit restored) | Return after about 6 business days | ≈ 4 days, repeatable | Φ₃ |
05Cite
BibTeX#
@inproceedings{malik2026creditwithoutdebit,
title = {Credit Without Debit: Formal Verification of Banking Payment Systems},
author = {Malik, Ali Hamza and Raza, Muhammad Taqi},
booktitle = {Proceedings of the 2026 ACM SIGSAC Conference on Computer and Communications Security (CCS '26)},
year = {2026},
address = {The Hague, Netherlands},
publisher = {ACM},
isbn = {979-8-4007-2871-6},
doi = {10.1145/3830454.3846614},
url = {https://doi.org/10.1145/3830454.3846614}
}
What is Credit-Without-Debit, in one sentence?
An attacker gets spendable credit at one institution while making the matching debit fail at another. We found three versions: creditBLEED advances funds, creditMINT leaves micro-deposits unrecovered, and creditRESET restores a card credit line before the bill payment settles. In each case, the attacker can use credit before the institutions resolve the failed debit.
As a normal customer, am I at risk, and what should I do?
Not from the three instances we reported: the institutions we tested have closed them. The underlying risk is broader. In creditBLEED, a bank started a debit using only routing and account numbers, and against an account without debit restrictions that debit moves real money. Review statements for ACH debits you did not authorize, and ask your bank whether it offers ACH debit blocks or filters.
Who lost money: the bank or the account holder?
In the attack design, the institution bears the loss. In our controlled creditBLEED test, Bank A reversed the credit, so it loses money only if the funds are withdrawn before the return arrives. Providers lose unrecovered micro-deposits, and card issuers expose restored credit to spending against failed payments. If the target account is not debit-restricted, creditBLEED instead transfers its money to the attacker. We tested only accounts we owned.
Which banks and payment providers were affected?
We anonymize them at their request. Our evaluation used researcher-owned accounts at two of the four largest U.S. ACH banks, three credit-card issuers, and three payment platforms. The paper uses labels such as Bank A and Bank B. These findings describe the institutions and workflows we tested. They do not establish that every bank is vulnerable.
Is this fixed?
Yes, for the three reported instances. creditBLEED and creditRESET were no longer viable by May 2025. creditMINT still worked at one provider through December 2025, but all affected providers had resolved it by January 2026. Follow-up testing confirmed the result. We did not receive detailed technical explanations of the changes, so we cannot describe their internal fixes.
Does an attacker need insider access or stolen credentials?
No. Our attacker has four ordinary customer capabilities: owning legitimate accounts, obtaining routing and account numbers without login credentials, initiating or receiving payments through consumer interfaces, and changing their own account settings through features such as holds or debit blocks. Banks do not collude with the attacker. Fraud detection and investigator intervention fall outside the model.
Did you break the banks' cryptography or hack their systems?
No. We used ordinary customer interfaces and accounts we controlled, without breaking encryption or bypassing login security. The weaknesses concern when institutions make credit spendable and when they check whether a debit can succeed. The paper identifies institutional choices permitted by the rules, rather than a cryptographic failure or a compromised banking server.
How much money is at stake?
We do not report an estimate of total losses. Each micro-deposit is small ($0.01–$0.99), but creditMINT issues more with each verification restart. creditBLEED exposes the amount the bank advances before the return, if withdrawn. creditRESET exposes the restored credit limit to new spending while the funding debit remains unsettled. These are exposure mechanisms, not measured losses across ACH.
What still needs to change in the rules?
The paper recommends stronger guarantees across institutions: recheck authorization before settlement, keep provisional credit unspendable until the debit passes access controls, and verify both credit acceptance and debit capability when linking accounts. A failed recovery should mean failed verification, and retries should not issue repeated credits. Standards should also clarify how authorization and settlement failures affect user-visible balances.
How is this different from card fraud or earlier payment research?
It concerns bank-transfer rules rather than card networks. To our knowledge, this is the first formal analysis of consumer ACH security. The attacker uses permitted customer actions to exploit the gap between spendable credit and debit settlement. Earlier research tested cards, wallets, and other payment platforms, and ACH fraud advisories describe return-code abuse. We use formal verification to identify the design-level causes.
What does model checking add beyond trying transfers manually?
It searches the modeled behaviors systematically, including intermediate states a manual test might miss. We translated the payment workflow into five communicating state machines and used nuXmv to check continuous authorization, balanced credits and debits, and settlement finality. The checker returned a violating sequence for each property. We then replayed those sequences on live accounts to validate the findings.
Does the model prove you found every possible attack?
No. Our analysis aims for soundness, not completeness: a reported violation holds under the model and its assumptions, but other violations may exist outside its bounds. We checked a finite instance and validated the three reported attacks on live accounts. This does not prove that every ACH workflow is secure or that every institution behaves alike.
Does this cover Same-Day ACH, wires, RTP, or card networks?
Partly. The model covers domestic consumer payments started online (WEB) and prearranged payments (PPD). It admits Same-Day ACH messages through an abstraction of processing windows, without modeling clocks within a day. Wires, RTP, and card networks are excluded. creditRESET involves paying a card bill through ACH. It does not establish an attack on the card network itself.
Was it ethical to test this on real banks?
Yes, by design. We confined every experiment to accounts under our control. No third-party account was affected, and we promptly informed the banks and providers bearing the financial losses. We disclosed through official channels on April 16, 2025, with reproduction steps, demonstrations, and technical reports. Institutions acknowledged the reports, and we supplied additional evidence for their internal review.
Can I see the model and reproduce the results?
Yes. The Zenodo archive contains the model source, property specifications, and scripts to reproduce the reported counterexamples. It excludes the copyrighted Nacha rules, which require registration or purchase. Raw account screenshots and transaction logs contain personally identifiable financial information and are withheld. The paper provides redacted screenshots of the observed state changes instead.