Industries / Finance and business services / Finance and banking / Fintech
Fraud
Fraud work in fintech is stopping criminals from opening accounts under invented or stolen identities, from taking over real customers' accounts and from tricking customers out of money, and handling customers' claims when it happens anyway.
5
Milestones achieved, from 2011 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2011
Random numbers aid fakes
On 25 June 2011 the US Social Security Administration began assigning Social Security numbers (SSNs) at random, to protect the numbers and to stretch the supply. That ended the link between a number's first three digits and the state that issued it, which lenders had used to spot made-up numbers. A 2019 Federal Reserve white paper says this made synthetic identities harder to detect; a synthetic identity is an invented person built from a mix of real and fake details, such as a real or unissued SSN under a false name. It cites the analytics firm ID Analytics' estimate that nearly 40% of synthetic identities use a randomized SSN.
LessonRemoving a pattern to protect people can also remove a signal that fraud checks depended on.
Credit
US Social Security Administration; the Federal Reserve, which described the effect
2023
Brazil pools fraud data
Under Joint Resolution No. 6 of 23 May 2023, Brazil's National Monetary Council and central bank required financial institutions, payment institutions and other firms the central bank authorizes to share data on suspected and attempted fraud with one another through an electronic system, from 1 November 2023. The central bank's follow-up rule of 4 October 2023 set the minimum scope: opening accounts with false documents, payments the customer did not consent to or was tricked into, accounts kept to receive fraudulent transfers (called "conta laranja", a mule account) and loans taken out in a customer's name without consent.
Credit
Conselho Monetário Nacional and Banco Central do Brasil
2024
Deepfakes at sign-up
On 13 November 2024 the US Financial Crimes Enforcement Network (FinCEN) warned that, "beginning in 2023 and continuing in 2024", it had seen more suspicious activity reports describing deepfake media, often identity documents altered or created with generative AI to get past identity checks. Criminals opened accounts this way and used them to receive and launder the proceeds of scams and other fraud. FinCEN's red flags include a photo that a reverse image search finds in an online gallery of AI-generated faces, and an applicant who claims "repeated technical glitches" to avoid a live video check; it points to phishing-resistant multifactor login and live verification checks as defenses.
Credit
FinCEN
2025
Cash App ordered to repay
On 16 January 2025 the US Consumer Financial Protection Bureau (CFPB) ordered Block, the operator of Cash App, to pay up to $120 million to harmed customers, at least $75 million, and a $55 million penalty. The CFPB found that Block did not fully investigate unauthorized transfers as federal law requires, told defrauded users to ask their banks to reverse transactions and then often denied the disputes, and printed a support number that led only to a recorded message; Block must now offer live customer service around the clock. On 10 April 2025 New York's Department of Financial Services fined Block a further $40 million over its anti-money-laundering program, finding that rapid growth in 2019 and 2020 had left a "severe transaction alert backlog".
LessonFor a money app, investigating claims and answering customers is part of fraud protection, and the duty cannot be passed to the customer's bank in the fine print.
Credit
CFPB; New York State Department of Financial Services
2025
Stolen logins buy stocks
From January to April 2025 criminals took over 6,380 accounts at Japanese online brokers and made 3,505 unauthorized trades, selling about ¥161.2 billion and buying about ¥143.7 billion of shares, by Financial Services Agency (FSA) figures quoted at the minister's press conference of 13 May 2025. They got the logins through fake websites posing as the brokers and used the accounts to buy "mainly low-grade stocks", in the FSA's words, rather than to withdraw money. On 2 May the Japan Securities Dealers Association (JSDA) and 10 brokers agreed to compensate customers in part, and in July 2025 the FSA asked that login protection that resists phishing be made mandatory by the end of June 2026.
LessonA fake website can collect a password and a one-time code alike, which is why the FSA asked for login methods that resist phishing.
Credit
Financial Services Agency, Japan; Japan Securities Dealers Association
Watch for
AI hunts mule accounts
A mule account is one used to pass on stolen money. MuleHunter.AI, a tool built by the Reserve Bank Innovation Hub, flags such accounts for Indian banks; in November 2025 RBI Governor Sanjay Malhotra said its adoption was "quite near" the intermediate goal of 20 banks, and the Free Press Journal reported that it was detecting about 20,000 mule accounts a month.
Achieved whenThe RBI or an outside party publishes how many flagged accounts were confirmed as mules and how many were false alarms.
Claimed by
Reserve Bank of India, November 2025
Shared fraud scores at sign-up
Plaid, which connects apps to people's bank accounts, sells Plaid Protect, a fraud score for sign-up, identity checks and bank linking built from signals across its network. Plaid says retro-analyses on its early customers' historical data show Protect "could have improved fraud detection by 40-59%" while asking 5% to 10% of users for extra checks, and would have prevented 52% of fraud dollar losses on average; it gives no sample size or method.
Achieved whenA customer or an outside party publishes measured results from live use, with the method.
Claimed by
Plaid, March 2026
Open
01
Regulation E disputes, benchmarked
A public test set of invented dispute cases for US app-based accounts, each with the customer's messages, account history and login records, labeled with the outcome and deadlines that the Regulation E error-resolution rule requires, and a scored comparison of AI agents that triage the disputes. Include the hard cases the CFPB's Cash App order describes: impersonation scams, customers sent to their bank, and claims that need provisional credit.
The briefCredit · unclaimed02
Synthetic identity rings, simulated
A generator of invented account and loan applications in which fraud rings build synthetic identities by reusing and varying details such as phone numbers, addresses, emails and ID numbers across many applications, with every ring labeled, and a scored comparison of open record-linkage tools at finding the rings. All details are made up, and ID numbers use a format that cannot belong to a real person.
The briefCredit · unclaimed
Open 01 · the brief
Regulation E disputes, benchmarked
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of invented dispute cases for US app-based accounts, each with the customer's messages, account history and login records, labeled with the outcome and deadlines that the Regulation E error-resolution rule requires, and a scored comparison of AI agents that triage the disputes. Include the hard cases the CFPB's Cash App order describes: impersonation scams, customers sent to their bank, and claims that need provisional credit.
Why it is still open
Regulation E requires a firm to investigate a claimed error within 10 business days, or within 45 days if it provisionally credits the money, and to report the result within 3 business days; the CFPB found in January 2025 that Cash App did not fully investigate unauthorized transfers. The nearest public benchmarks test regulatory compliance in AI tool use in general (LogiSafetyBench, 240 tasks, January 2026) or under Chinese financial rules (FinGuard, May 2026). This research found none for dispute handling.
Built with what exists
The rule's text on the CFPB's website; a general AI model such as Claude Fable 5.2 (see the Note "Three working devices") to draft cases for people to correct; as agents under test, Gemini Decision 1, which returns probabilities for fixed-answer questions (see the Note "A thousand questions, one pass"), and an open model such as Mistral Large 4 (see the Note "A trillion open parameters"). No real accounts or money are involved.
First step
Write ten cases from the patterns in the CFPB's Cash App order and label each with its 10-day and 45-day dates.
Done looks like
At least 300 cases, each labeled by two people who agree or record how they settled a difference, released under an open license with a scoring script, and a table of at least three agents showing correct outcomes and missed deadlines. The finish line is an agent that reaches the labeled outcome on at least 95% of cases and misses no deadline.
Open 02 · the brief
Synthetic identity rings, simulated
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A generator of invented account and loan applications in which fraud rings build synthetic identities by reusing and varying details such as phone numbers, addresses, emails and ID numbers across many applications, with every ring labeled, and a scored comparison of open record-linkage tools at finding the rings. All details are made up, and ID numbers use a format that cannot belong to a real person.
Why it is still open
The Federal Reserve calls synthetic identity fraud "reported to be the fastest-growing type of financial crime in the United States". Feedzai's Bank Account Fraud dataset, released at NeurIPS 2022 with six million synthetic account applications, generates each application on its own, "with no links to other instances", and IDNet covers forged document images, not links between applications. SynthWorld (July 2026) publishes small synthetic identity graphs with answer keys for testing how tools match records about the same person, but not account applications with planted fraud rings. This research found no public dataset of applications linked by shared details with labeled rings.
fedpaymentsimprovement.orgraw.githubusercontent.comarxiv.orghuggingface.co
Built with what exists
Splink, the UK Ministry of Justice's open record-linkage package (MIT license), which links "a million records on a laptop in around a minute" by its own account; a graph library such as NetworkX to find clusters; the Bank Account Fraud dataset's feature list as a template for application fields; SynthWorld's rules for making every record plainly fake, such as email addresses on the reserved example.test domain and ID numbers with a SYN- prefix and invalid check digits.
First step
Write a script that makes 10,000 applications with 20 planted rings, and run Splink on it.
Done looks like
A generator and a released set of at least 1 million applications with at least 1,000 labeled rings, and a table scoring at least two tools on the precision and recall of ring membership; another team can regenerate the set from a fixed seed and get the same scores.