A data error at FNB became a privacy breach. Hygiene and security are one problem.
South Africa’s Information Regulator fined FNB R1 million after customers wrongly flagged as bank employees could open other people’s home loan applications. The access rule worked. The data it read was wrong, two warnings were closed as one-off tickets, and some records were kept with no stated reason. That is a data hygiene failure and a security failure at the same time, and AI agents will read the same flags.
What happened
On 26 October 2026, South Africa’s Information Regulator announced a R1 million fine against First National Bank under the Protection of Personal Information Act, together with an enforcement notice. The regulator found weaknesses in the security safeguards on FNB’s digital home loan platform, and held that FNB committed an offence over the exposure of account numbers.
TechCentral reported the detail from the enforcement notice, dated 21 October. A misconfigured setting in the FNB app let 696 customers see the home loan applications of 88 other people. Some account holders had been wrongly classified as FNB employees. In the app, that flag showed them a list of names and ID numbers of people with applications in progress, and a click opened the full application: contact details, address, income and, for nine people, account numbers.
A customer reported the problem on 30 March 2022, and another on 3 May 2022. Each time, FNB removed that person’s access, treated the report as an isolated incident and did not tell the regulator. The regulator became involved after a complaint in September 2022, and called the bank’s handling “lackadaisical”.
The notice went beyond the incident. FNB’s retention schedule allowed some records, including cancelled employee tax certificates and securities registers, to be kept indefinitely with no stated legal basis, and the bank must now have them destroyed and certified. The regulator also found that FNB could not rely on newspapers as a public record for the information it collects to screen politically exposed persons.
My view
This case is small in numbers and large in what it shows. Nobody broke in. An access rule did exactly what it was told, on data that was wrong.
The breach was a classification error
The app granted access based on one attribute: is this person an employee? That attribute lived in customer and HR data, and for some account holders it was wrong. A data quality defect in a field most people would call reference data became, the moment the app read it, an access control failure.
Every access rule is a question asked of data. Who is an employee, which branch, which role, which customer segment, which jurisdiction. Security teams review the rule. Far fewer firms treat the fields the rule reads as critical data, with an owner, a definition, quality tests and a record of who changed them. In my experience that is the gap: the logic is reviewed, and the data it depends on is assumed.
Two warnings, closed as tickets
The handling matters as much as the defect. Twice, a customer said in effect “I can see other people’s applications.” Twice, the bank fixed that one customer’s access. The question that would have found the problem is a data question: who else carries the same wrong flag, and what can they see? A security incident handled one user at a time, without looking at the population, stays open while the tickets close.
Data kept without a reason is exposure without a purpose
The retention finding is the same problem from the other side. A record kept past its purpose earns nothing and can still be lost. New York’s financial regulator made the point in April, when it settled with Delta Dental for $2.25 million over the 2023 MOVEit breach. According to the consent order, the companies had extended file retention settings past the 30-day default, and in some cases switched them off, with no written process for approving those changes. Some of the files the attackers took had been kept longer than 30 days. Retention is usually filed under records management. In practice it sets how much a breach costs.
Agents will read the same flags
This is where it gets more urgent. An AI agent acting for a user inherits whatever that user’s data says they may see. If the role data is wrong, the agent is wrong too, and it reads faster and more widely than a person browsing an app. An assistant asked to “summarize the home loans in progress” on behalf of a wrongly flagged customer would not stop at one click. Before you give agents a user’s access, check the data that grants it.
Why it keeps happening
The work that prevents cases like this has no launch date and no revenue line: cleaning role and party data, reviewing entitlements against real populations, deleting what you no longer need. It is the first work deferred when budgets are tight, and it stays deferred until a regulator orders it. FNB now has to document, destroy and certify on a 90-day clock. Doing it before the notice would have been cheaper.
Why it matters
For the industry. Data hygiene and security are usually run by different teams with different budgets. Cases like this show they are one control. The fields that decide access are critical data elements, the records you keep set the size of any breach, and incident handling has to look at the population, not just the person who reported it. As agents take on user access, a wrong flag stops being a one-person exposure and becomes a machine-speed one.
How it connects to our work. Data Product Certification gives critical data an owner, a contract and tests, including the fields access rules depend on. The Reference Implementation enforces access on every approved path from labels and roles you can inspect. Agent Onboarding bounds what an agent can reach regardless of what a user’s record says. AI Governance Evidence covers the incident records a regulator will ask for. Six steps I’d take first:
- List the fields your access rules read. Role, employee flag, customer type, branch, jurisdiction. Give each an owner, a definition and a quality test.
- Test access against real populations. Ask who can actually see what, not only whether the rule is written correctly.
- Look past the reporter. When someone reports seeing another person’s data, find everyone with the same attribute before you close the ticket.
- Give every retention setting an owner and a written reason. Alert on any change that extends or disables it.
- Delete on schedule and keep the proof. A destruction certificate is cheaper before the regulator asks for one.
- Check the data before an agent inherits it. An agent should never see more than a correct record would allow.
This is one of three Viewpoints on why security, data hygiene and resilience are one problem. The others: Microsoft’s outage started with a fix from July and Most AI data leaks run through approved tools.
Sources
- Regulator hits FNB with R1-million Popia fine, TechCentral, 27 October 2026, reporting the enforcement notice dated 21 October 2026.
- Information Regulator media statements, Information Regulator (South Africa), statement of 26 October 2026.
- Incident response and data retention protocols at center of NYDFS enforcement action against Delta Dental, WilmerHale, 1 June 2026, summarizing the consent order; Delta Dental agrees to $2.25 million settlement with NYDFS over MOVEit data breach response, Hunton Andrews Kurth.
Linked sources are the record. Check them for the full reporting.