South Africa Joint Standard 2 of 2024: cyber resilience and AI in finance

In a Johannesburg bank, a risk analyst is preparing a note for the credit committee. She pastes an extract of the loan book into an AI assistant, with client names, incomes and arrears, to get a clean summary. On the same floor, a developer asks a coding tool to fix a script that contains an access key to the client database.

A few years ago, these actions were mainly a matter of POPIA and internal compliance. Since 1 June 2025, they also fall under a prudential standard: Joint Standard 2 of 2024 on Cybersecurity and Cyber Resilience, issued jointly by the FSCA and the Prudential Authority. For South African banks, insurers and other financial institutions, AI use is now a supervisory question.

A standard shared by the FSCA and the Prudential Authority

Joint Standard 2 of 2024 was published on 17 May 2024 by the Financial Sector Conduct Authority and the Prudential Authority, which sits within the South African Reserve Bank. It took effect on 1 June 2025. It complements Joint Standard 1 of 2023 on IT governance and risk management, in force since 15 November 2024.

Its scope is broad. It covers in particular:

• banks and mutual banks;

• insurers;

• market infrastructures: exchanges, central securities depositories, clearing houses;

• registered pension funds and certain fund administrators;

• over-the-counter derivative providers and registered credit rating agencies.

What the standard requires for sensitive data

The text does not mention generative AI as such. But several of its requirements apply directly to what goes into a prompt. Institutions must keep an inventory of their information assets, limit access to information to authorised users and devices, and prevent unauthorised use of sensitive data. They must also deploy endpoint protection and regularly test how well their controls work.

A consumer AI assistant used without oversight contradicts almost every one of these points. Data leaves the perimeter for a service that appears in no inventory, with no usable trail. And classic network filtering tools struggle to see what is happening, as we explain in why email and network DLP fails on ChatGPT.

Reporting an incident within 24 hours

Institutions must notify the responsible authority of any cyber or IT incident classified as material. Since 1 September 2026, a common reporting template applies, set by Joint Notice 2 of 2026. The first notification must be sent within 24 hours of the incident being classified as material, weekends and public holidays included. An update is expected within 14 calendar days.

Banks and insurers report to the Prudential Authority. Other covered institutions report to the FSCA. At the same time, POPIA requires notifying the Information Regulator and the people affected when personal data is compromised. A leak of client data through an AI tool can therefore trigger two reporting tracks at once.

AI is spreading fast in South African finance

At the end of 2025, the FSCA and the Prudential Authority published a study on AI in the financial sector, based on a survey run between October and December 2024. It drew about 2,100 responses. 52% of banks and 50% of payment providers said they already used AI, compared with 8% of insurers.

The risk respondents cited most often was personal data protection under POPIA. Cybersecurity, model risk and risks specific to generative AI, such as inaccurate outputs, came next. Regulators therefore have a clear view of the issue, and it would be unwise to assume they will overlook it in coming reviews.

The uses that cause problems

In financial institutions, a few situations keep coming up:

1. Credit analysis: file extracts with South African ID numbers, incomes and payment histories.

2. Claims: assessor reports and medical records summarised to speed up a settlement.

3. Customer service: complaints pasted in with account numbers and contact details.

4. Development: scripts and configuration files holding secrets, a risk covered in our article on how CLI AI tools leak your secrets.

For the specific case of the 13-digit ID number, we published a dedicated guide: detecting and anonymising South African ID numbers before AI.

How ZeusLock helps financial institutions

ZeusLock runs on employees' devices and browsers. Before a prompt or file reaches ChatGPT, Copilot, Gemini or Claude, it detects ID numbers, account numbers, cards, API keys and other sensitive data. Depending on the policy, it anonymises them while keeping their format, warns the user or blocks the request. Every event is logged and timestamped. These logs help show that the use of sensitive data is under control and help qualify an incident quickly, which matters once the 24-hour clock starts. The full path of a leak is described in our anatomy of an AI data leak.

Summary

With Joint Standard 2 of 2024, cybersecurity at South African financial institutions has become a precise prudential obligation with short reporting deadlines. Generative AI, already used by half of the banks, opens new leak paths that the standard requires firms to control. Managing what goes into prompts is one of the most concrete ways to do it.

Do you work at a bank, insurer or asset manager in South Africa? Book a ZeusLock demo and see how to govern AI use.