Neobanks Ship the Future on a Shaky Foundation as AI Agents Arrive at the Bank Door

The gap is in the basics, not the features

Independent neobanks look modern on the surface. In Ostorlab's study, 37.9% of their apps integrate a passkey API, against 4.6% of traditional-bank apps, and 48.3% use app or device attestation, against 19.9% at traditional banks. The same study finds the foundations weaker: 79.3% of independent neobank apps carry at least one high or critical security finding, against 51.2% of traditional-bank apps, and 51.5% of all apps in the study carry a high or critical finding. m:…3869

The most common failures are the ones that hand an attacker the keys. The study found 48.3% of independent neobank apps embed credentials beyond Google API keys, against 19.8% of traditional-bank apps, and independent neobanks ship embedded credentials more than twice as often as traditional banks. High or critical OAuth account-takeover findings appear in 24.1% of independent neobank apps, against 6.1% at traditional banks, and 41.1% of all apps in the study ship a third-party component with a published vulnerability. m:…3869

The stakes are already visible in scam losses

Mobile banking accounted for 52% of UK authorized push-payment scam losses in 2025, according to the study, which puts the app-security gap in the middle of the fraud problem rather than beside it. Only 22.3% of Android banking apps detect remote-control tools used to commit fraud, and 79.8% of apps show no app or device attestation, so the tools that would prove which device is really talking to the bank are missing in most of the market. m:…3869

Japan shows what happens when attackers move faster than patching. Japan disclosed 233 security incidents in October 2026, with approximately 185 million personal information records exposed in that month. JPCERT/CC confirmed that attackers used AI to scan for vulnerabilities and conduct attacks through APIs, and the Japanese government announced a 10 billion yen emergency cybersecurity package in response. m:…9486

The breaches behind the total were large and concentrated in everyday services. Arara confirmed an intrusion affecting more than 81 million email addresses, Nitori disclosed that unauthorized access may have exposed up to 37.62 million customer records, and Ymir Link disclosed an intrusion into its Cuenote FC service involving 47.9 million records. Japan plans to seek earlier access to information about undisclosed software vulnerabilities from US AI developers, a recognition that the exposure starts upstream, in software the whole country runs on: PHP runs on approximately 92% of websites hosted in Japan, and WordPress powers 88% of Japan-hosted websites with an identifiable CMS. m:…9486

Agents are the next visitor at the door

Banks are now being visited by software that is neither a customer nor an ordinary bot. Feedzai reports that seven in eight agents arrive with no name or owner, and that traditional login screens and bot detection cannot reliably determine whether an agent is legitimate. The company says JPMorgan Chase is blocking or restricting agent access to customer accounts, a sign that at least one large bank would rather shut the door than guess who is knocking. m:…6204

Feedzai says its answer is to check the visitor rather than the password, using four signal layers: device intelligence, behavioral biometrics, network and threat intelligence, and endpoint security. The company plans to release agent fingerprinting in Q1 2027 and agent behavioral intent analysis in Q2 2027. Until tools like those are in place, a bank that cannot tell an agent from a customer is deciding, case by case, whether to let software act on a customer's money. m:…6204