BRENNEN

Industries / Finance and business services / Finance and banking / Credit unions

Fraud

Theft from credit unions and from their members, by insiders, outside criminals and scammers, and the checks and shared warnings that catch it.

5

Milestones achieved, from 2014 to 2023.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2014

    Straw borrowers drain a SKOK

    In Poland, cooperative savings and credit unions are called SKOKs. Prosecutors in Szczecin say the Polish Financial Supervision Authority (KNF) knew by September 2012 that SKOK Wołomin was lending on a large scale to straw borrowers, people put forward to sign for loans they could not repay, and that its own inspection confirmed this in spring 2013. KNF put in an administrator only on 4 November 2014, after the credit union's president and vice-president were arrested; the prosecutors put the loss from the delay at about 1.7 billion złoty, an average of about 14 million złoty a day. The administrator found 79.8% of the loan book overdue, and KNF suspended the credit union on 11 December 2014, which by KNF's notice opened the way to payouts from the Bank Guarantee Fund (BFG), Poland's deposit insurer.

    Lessononce a supervisor has evidence of fraud, the losses keep growing until someone takes control of the institution.

    gov.plintranet.knf.gov.pl

    Credit

    KNF, which suspended it; the Regional Prosecutor's Office in Szczecin, which set out the findings

  2. 2019

    Twenty years undetected

    NCUA liquidated C B S Employees Federal Credit Union on 29 March 2019 after examiners found about $42.2 million in cash missing. Its chief executive had embezzled alone for about 20 years. He wrote official credit union checks to payees of his choice, paid his own bills with them, and hid the gap by understating members' share certificates, which are fixed-term savings, in the financial statements. He held the check stock, had "super-user" access that let him alter the accounting records, and was solely responsible for financial reporting. NCUA's Inspector General put the loss to the share insurance fund at $39.5 million and found that examiners had not treated the credit union's supervisory committee audits and member account verifications as unacceptable when they should have.

    Lessonone person who writes the checks, keeps the books and reports the results can hide theft for decades; comparing the ledger with balances confirmed by members is the check that exposes it.

    ncua.gov

    Credit

    NCUA Office of Inspector General, with the auditors Moss Adams LLP

  3. 2019

    An insider copies member data

    On 27 May 2019 the Fédération des caisses Desjardins du Québec, the federation of Desjardins credit unions (caisses) in Quebec, reported a breach that ultimately affected close to 9.7 million people. An employee had copied names, dates of birth, social insurance numbers, addresses and transaction histories for at least 26 months, from a marketing department shared drive to his work computer and then to USB keys. Desjardins did not detect it; the police told it. The Office of the Privacy Commissioner of Canada found in December 2020 that Desjardins had broken federal privacy law on accountability, retention and security safeguards, including that its systems did not block USB keys and it had no active monitoring of data access.

    Lessoninsider theft of data is caught by limiting and watching access to member records, not by the perimeter defenses built against outsiders.

    priv.gc.ca

    Credit

    Office of the Privacy Commissioner of Canada, with Quebec's Commission d'accès à l'information

  4. 2019

    Elder theft, counted

    Banks and credit unions file suspicious activity reports (SARs) with FinCEN, the US Treasury's financial crimes unit. In what it calls the first public analysis of SARs on elder financial exploitation since FinCEN added that category in 2013, the Consumer Financial Protection Bureau (CFPB) found that such reports quadrupled from 2013 to 2017, to 63,500 in 2017, covering $1.7 billion in suspicious activity. Where an older adult lost money, the average loss was $34,200. Fewer than one-third of the reports said the filer had also told adult protective services, the police or another authority.

    Lessona report filed with the Treasury does not reach the local agencies that can help a victim unless someone also calls them.

    consumerfinance.govfiles.consumerfinance.gov

    Credit

    CFPB Office for Older Americans, with FinCEN data

  5. 2023

    Cyber incidents within 72 hours

    Since 1 September 2023 NCUA has required federally insured credit unions to notify it of a reportable cyber incident as soon as possible, and within 72 hours. In the first year, to 31 August 2024, credit unions filed 1,072 reports, and 742 of them, nearly 7 in 10, involved a third-party vendor, by NCUA staff's count. Staff named two trends: outages caused by ransomware attacks and outages caused by service providers.

    Lessonmost incidents a credit union reports start at a company it pays, not inside its own walls.

    ncua.govncua.gov

    Credit

    NCUA Board, which adopted the rule in February 2023

Watch for

  1. Name checks before paying

    Confirmation of Payee checks whether the account name a customer types matches the name the receiving bank holds, before a transfer is sent. Australian Payments Plus (AP+), which runs the service, says 83 institutions offer it, reaching over 143 million accounts, and that it has been used more than 100 million times since its July 2025 launch. The Customer Owned Banking Association (COBA) says mutual banks and credit unions worked to put it in, and one unnamed bank reports over 450,000 payments abandoned after a "no match" result. None of the figures has been checked by anyone outside the industry.

    Achieved whena regulator or other independent body publishes scam losses that separate payments checked by Confirmation of Payee from those that were not, and shows lower losses on checked payments.

    auspayplus.com.au

    Claimed by

    AP+, the Australian Banking Association and COBA, March 2026

Open

  1. 01

    Plant thefts in test books

    A public set of synthetic books for imaginary small credit unions: member share and loan accounts, the general ledger, the official check register, file maintenance logs (the record of who changed which account settings) and quarterly reports. Some of the credit unions carry a planted insider scheme modeled on published cases, such as official checks hidden by understating share certificates, as at C B S Employees, or loans to straw borrowers, as at SKOK Wołomin, and others carry none. The set tests whether the standard audit steps of a credit union's supervisory committee, written as scripts, and AI agents given the same files find the schemes.

    The brief
    Credit · unclaimed
  2. 02

    Spot a coached member

    A test set of staged conversations between members and credit union staff, by phone and at the counter. In some, the member is being coached by a scammer to move money, for example by a fake "fraud investigator" or a romance contact; in others, members make the same kind of request for genuine reasons. Actors play both sides from written scripts, with no real member data. The set measures whether staff-assist tools, which transcribe a call and classify it, raise a flag, and how often they flag a genuine request.

    The brief
    Credit · unclaimed

Open 01 · the brief

Plant thefts in test books

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of synthetic books for imaginary small credit unions: member share and loan accounts, the general ledger, the official check register, file maintenance logs (the record of who changed which account settings) and quarterly reports. Some of the credit unions carry a planted insider scheme modeled on published cases, such as official checks hidden by understating share certificates, as at C B S Employees, or loans to straw borrowers, as at SKOK Wołomin, and others carry none. The set tests whether the standard audit steps of a credit union's supervisory committee, written as scripts, and AI agents given the same files find the schemes.

Why it is still open

NCUA's own review of 16 failure reviews from 2015 to 2021 found fraud in 13, and says "almost every fraud instance was due to insider fraud"; it publishes summary counts only. Tests of fraud detection in accounting entries use company books: Schreyer and colleagues' 2017 work with autoencoders, a kind of neural network, used real journal entries that were not released, and a December 2025 study by Md Abdul Kadir and colleagues at DFKI tested language models such as LLaMA and Gemma on synthetic and anonymized company ledgers. Open generators of company general ledgers with injected anomalies exist on GitHub. A search found none that models a credit union's member share accounts, official check register and file maintenance logs, where the published insider schemes happened.

ncua.govarxiv.orgarxiv.org

Built with what exists

Apache Fineract, an open-source core banking system from the Apache Software Foundation, to hold the member and loan records; Python with the Faker library to generate members; an AI agent such as Claude or Gemini with read-only access to the files and the audit steps (see the Note "Seven more flagged", on an AI model finding errors in production records that two careful human reviewers had missed).

fineract.apache.org

  1. First step

    Generate one credit union with 500 members, plant the C B S scheme, and check whether reconciling the member share ledger with the general ledger finds the gap.

  2. Done looks like

    100 synthetic credit unions of 300 to 3,000 members, 40 with one of 8 planted scheme types; for the scripted audit steps and for at least two AI agents, the share of schemes found and the number of false alarms, with 95% confidence intervals; the generator, data and results published under an open license.

Open 02 · the brief

Spot a coached member

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A test set of staged conversations between members and credit union staff, by phone and at the counter. In some, the member is being coached by a scammer to move money, for example by a fake "fraud investigator" or a romance contact; in others, members make the same kind of request for genuine reasons. Actors play both sides from written scripts, with no real member data. The set measures whether staff-assist tools, which transcribe a call and classify it, raise a flag, and how often they flag a genuine request.

Why it is still open

Phone scam datasets record the call between the scammer and the victim. Shen and colleagues, reviewing detection with large language models in 2024, point to biased datasets and low recall, and a 2026 Korean system was trained on a corpus of 185,300 scam cases. Neither covers the later moment when a coached member asks their own credit union to move the money. The CFPB found that most elder exploitation reports were not passed on to local authorities, which makes that moment the last chance to stop a loss.

arxiv.orgarxiv.orgfiles.consumerfinance.gov

Built with what exists

volunteer actors; scripts drawn from public scam warnings; a speech-to-text model such as ElevenLabs Scribe v3 (see the Note "Fewer word errors"); a model that answers many fixed questions about a text in one pass, such as Google's Gemini Decisions API (see the Note "A thousand questions, one pass").

  1. First step

    Write 20 paired scripts, one coached and one genuine for the same request, record them with two actors, and run them through one transcription and classification pipeline.

  2. Done looks like

    300 staged conversations, half with signs of coaching, labeled by two experienced fraud staff with their agreement reported; the recall and false-alarm rate of at least three tools, with 95% confidence intervals; scripts, audio and labels published.