Customers who give an AI agent their banking login could carry the loss if something goes wrong, Absa has warned, as South African banks join their US rivals in keeping outside AI agents away from customer accounts.
FNB, Discovery Bank and Capitec say they don’t allow third-party AI agents into customer accounts. Absa strongly discourages customers from handing their credentials to one, and warns that those who do may be held liable for any losses. FNB, meanwhile, is building a way for approved AI assistants to connect in future.
TechCentral asked the banks about the risks after Axios reported last week that major US banks had detected unauthorised AI agents in customer accounts, some of which appeared to act independently of users’ instructions. Reuters followed up the report. Morgan Stanley has begun blocking third-party agents, Axios said, though a spokesman said it had not seen “rogue AI agents accessing our platforms”.
JPMorgan Chase CEO Jamie Dimon said on the bank’s third-quarter earnings call on 13 October that it was “not going to allow just any outside agent” in, because “that’s just too risky for clients and for JPMorgan”.
Standard Bank and Nedbank had not responded to TechCentral’s questions by the time of publication.
“FNB does not currently allow third-party AI agents to access customer accounts or make transactions for customers,” said Dané Liebenberg, AI executive for retail and business banking at FNB. “We believe AI assistants will become an important way for customers to bank, but access has to be safe, governed and controlled by the customer.”
FNB is “building a secure framework for approved AI assistants to connect to FNB in the future”, Liebenberg said. Customers would give “explicit consent for every action”, identity checks and authentication would “always happen on FNB’s secure platforms” and every interaction would be “monitored and recorded”.
‘No autonomy without control’
The bank will start with information services and expand “only once the right protections are in place”. It is also investing in its own AI tools, including its digital assistant, NAVi. “Our principle is simple: no autonomy without control,” Liebenberg said.
Capitec was the most categorical. It said it does not permit third-party AI agents to access client accounts, and “that includes screen-scraping tools”.
“A client’s account can only be accessed by the client, through Capitec’s own secure channels. It is currently not possible for an agent to access our accounts,” the bank said. “At Capitec, we would not allow an ungoverned agent in our clients’ accounts.”
That may hold for agents acting on their own. But as Absa points out, an agent given a customer’s login details can sign in as that customer. Capitec did not say whether it had seen clients share their login details with agents, which it tells them not to do.
Capitec said it is building its own tools “in a way that is safe, governed and useful for our clients”, but did not say whether it would let approved outside agents connect. The bank said it is “not yet providing AI solutions directly to our clients”, though group CIO Andrew Baker wrote on 7 October that it was rolling out Scam Checker, “a free AI tool inside the Capitec app” for checking suspicious links, text, screenshots and images, over the following weeks.
“At Absa, we do not use AI agents to access customer accounts, and we strongly discourage customers from sharing their banking credentials with any third-party AI agents,” said Jochem Melis, CIO of Absa’s engineering, services and platforms office.
Melis was explicit that an agent armed with a customer’s login details could act on the account. “These agents do not independently gain access to customer accounts. Access is only possible when an authorised customer provides the necessary credentials or permissions,” he said.
“The primary concern is not technology itself, but the sharing of credentials. From a security perspective, providing banking credentials to an AI agent is no different from sharing them with any other third party.”
Unlike its US counterparts, Absa doesn’t see agents as a new threat. They “do not materially alter the bank’s current risk profile, as the underlying security risks are already addressed through established safeguards”, Melis said, adding that its fraud systems are designed to spot “suspicious patterns, including those consistent with unauthorised agent access attempts”.
On building its own agent tools or a route for outside agents, Absa said: “We’re exploring emerging standards for secure agent interaction and will share more as this evolves.”
Discovery Bank said it “does not currently allow third-party AI agents to access client accounts or initiate transactions”. It is instead building its own agentic tools through Discovery AI, which lets clients prepare a payment in the Discovery Bank app using natural language or an image of an invoice. The client must review and approve the payment before it is processed. Any loss would be “investigated on its individual facts, including how access was obtained and whether the transaction was authenticated and authorised”, the bank said.
Who pays?
All four banks said they would look at losses case by case.
“Where a customer voluntarily shares such information and a financial loss subsequently occurs, the customer may be held liable,” Melis said. Each case would be assessed under the Absa digital fraud warranty, which he said “is designed to protect customers who transact responsibly and adhere to the bank’s security requirements”.
Capitec said clients should not share their Pin, password or one-time Pin “with any third party, including AI tools or automated agents”. Where a loss follows, “we assess the circumstances of each case in accordance with our terms and applicable law”.
FNB said it would “assess any reported loss case by case, in line with our terms and conditions and the applicable regulations”. “We strongly urge customers never to share their banking login details, Pins or one-time Pins with anyone, including AI tools or apps,” Liebenberg said.
In the US, such transactions may not count as unauthorised. Federal rules exclude transfers by someone a consumer has given account access to, though the Consumer Financial Protection Bureau says account agreements cannot cut back consumers’ rights under the rules, and customers tricked into sharing their details remain protected.
Banks are also preparing for agents they do trust. Visa has enrolled South African banks in its Agentic Ready programme to prepare them for AI shopping agents. And on 15 October, the Financial Stability Board, which co-ordinates financial regulation for the G20, finalised its guidance on AI in financial institutions, following a June draft that urged banks to rein in AI agents. Among its sound practices is “appropriate and effective human oversight”. — © 2026 NewsCentral Media
