Brazil banks and AI: Pix fraud, social engineering and data leaks

Since the Central Bank of Brazil launched it in November 2020, Pix has changed the way Brazilians pay. A transfer goes through in seconds, at any hour, using a simple key: a CPF, a phone number or an email address. That speed has also changed fraud. When money moves in seconds, an attacker no longer needs to break a system. Convincing a person is enough.

In banks and fintechs, that person is often an employee. And that employee now uses generative AI assistants to work faster, sometimes with customer data right in front of them.

Pix has shifted fraud towards people

Fake call centres, fake advisers on messaging apps, fake technical support: the most common scams rely on social engineering. They target customers, but also internal teams. A rushed call centre agent, an analyst working through a queue of disputes or an IT contractor can become the way in.

The C&M Software case in 2025 showed this in a brutal way. According to the São Paulo civil police, an employee of a provider that connects institutions to the Central Bank's system admitted selling his credentials to criminals. They then diverted funds from the reserve accounts of several institutions. It was not a sophisticated technical flaw that opened the door, but a person.

What CMN Resolution 4,893/2021 already requires

CMN Resolution No. 4,893 of 26 February 2021 sets the cybersecurity policy for institutions authorised by the Central Bank, along with the rules for using cloud services. Payment institutions follow equivalent Central Bank rules. The text, updated at the end of 2025 by CMN Resolution No. 5,274, requires in particular:

• mechanisms to prevent information leakage, alongside authentication, encryption and traceability;

• recording and analysis of relevant incidents, including those reported by service providers;

• training programmes and periodic staff assessments to spread a cybersecurity culture;

• informing customers about the precautions to take with financial products.

In other words, leak prevention and the human factor are already regulatory obligations. The 2025 update adds multi-factor authentication for administrative access to Pix environments and an annual penetration test.

Risky habits with AI, day to day

In a bank or fintech, there are many uses of generative AI, and most are legitimate. The risk appears when the pasted content holds customer data. A few typical examples:

• an agent pastes the history of a Pix dispute, with CPF, name and amount, to draft a reply to the customer;

• a fraud analyst asks an assistant to spot a pattern in a transaction extract;

• a developer shares application logs containing Pix keys or account numbers to debug an issue;

• a relationship manager has a small business credit file summarised before a meeting.

In each case, data covered by bank secrecy (Complementary Law No. 105 of 2001) and by the LGPD (Law No. 13,709 of 2018) leaves the institution's controlled environment. We listed the identifiers to watch, starting with the CPF, in our article on the LGPD and detecting CPF numbers in AI tools.

Why training alone is not enough

Brazilian banks already train their staff, and Resolution 4,893 requires it. But an annual course carries little weight against a queue of waiting customers or a processing target. The employee knows sensitive data should not be shared. At the moment of pasting, they mostly think about finishing the task. We explain this in our article on why AI security training doesn't work.

Behavioural security therefore means acting at the right moment: a clear warning when pasting, automatic anonymisation where possible, a block when the risk is too high. The employee keeps the tool, and the institution keeps control.

The ANPD and the banking regulator look the same way

The National Data Protection Agency (ANPD), now a regulatory agency, enforces the LGPD, which requires security measures and the notification of incidents that may cause significant risk or harm to individuals. The Central Bank, for its part, expects institutions to record and handle their incidents. A leak of customer data to an external AI service can therefore concern both authorities at once.

An action plan for risk and compliance teams

1. Map AI use by business line: customer service, fraud, credit, development.

2. Classify the most sensitive data: CPF, CNPJ, Pix keys, account and card numbers, credit scores.

3. Anonymise at the source while keeping the format, so AI stays useful. Our article on format preserving anonymization shows how.

4. Link alerts to the incident register required by Resolution 4,893.

5. Roll out in stages, starting with the most exposed teams, as described in our 90-day AI DLP rollout playbook.

Summary

In the Pix era, fraud targets people more than systems. Brazilian banks and fintechs are already required to prevent information leaks and train their teams. Generative AI adds a new leak channel, fed by ordinary habits. A control at the moment of pasting protects customers without slowing teams down.

ZeusLock detects CPF numbers, Pix keys, account and card numbers before they reach an AI assistant, and records each incident for your compliance reports. Book a ZeusLock demo to protect your teams right where they work.