Successful Authentication Leaves Other Questions Open
Feedzai says Digital Trust combines endpoint, device, behavioural and network information before login and throughout the customer journey. Persistent device fingerprinting is intended to recognise returning activity, while a risk score draws on shared intelligence that may reveal patterns an individual institution cannot see alone. Cryptographic device binding adds a stronger connection between a recognised device, its customer and the bank.
“Personal and malicious agents can look very similar,” said Pedro Barata, Feedzai’s Chief Product Officer. “Our latest advancements in Digital Trust help banks tell them apart, blocking unwanted agents and trusting the ones their customers choose to use.” The difficulty embedded in that statement is important. Automation is a description of how a session operates; legitimacy depends on authority, intent and the particular action being taken.
Recognising an Agent Is Not the Same as Granting Permission
Feedzai describes an approach that combines agents’ declared characteristics with device and behavioural signals indicating automated activity. Once a relationship is recognised, banks can reduce repeat challenges where their policies permit. The institution retains control over allowed activity and how customer consent is handled. The company says it will expand these capabilities further in the coming months, making the release part of an ongoing development programme.
The consent question remains distinct from device recognition. A recognised device may provide useful evidence about continuity, but a customer’s approval for one task does not necessarily extend to another. In practice, institutions will need policies defining what an agent may do, how that authority is recorded and when a change in behaviour requires renewed confirmation. These are governance choices alongside the technology’s detection functions.
The challenge resembles the broader deployment issues highlighted by research into financial firms’ AI orchestration. As agents act across systems, their usefulness depends on reliable data and controlled permissions. Banks will need to assess whether an access decision remains appropriate throughout a workflow, particularly when a session moves from retrieving information to changing account details or initiating a payment.
Authentication Costs Provide a Nearer-Term Business Case
The expansion also targets existing human authentication problems. Feedzai says binding a trusted device to the customer can support a possession factor and reduce reliance on SMS one-time passcodes. It also describes checks for device compromise and social engineering around passkey use, account recovery, new-device enrolment and call-centre authentication linked to a digital session. Those claims concern the wider access process, rather than a claim that passkey cryptography itself has been broken.
That wider perspective matters because criminals can target routes around the principal login mechanism. Research into malware targeting banking applications illustrates why the endpoint remains relevant even when credentials are valid. A stronger authentication method does not remove the need to examine the device, account recovery process and surrounding behaviour for signs that control has changed.
Feedzai reports 41% fewer SMS passcodes and 6,000 fewer unnecessary authentication challenges per million logins compared with competing technology. The latter represents 0.6% of all logins, rather than a disclosed percentage reduction in the existing challenge rate. The results suggest a potential cost and convenience benefit, but the announcement does not provide enough detail about the comparison population and testing conditions to establish what another bank would achieve.
The Fraud Figures Need Their Original Denominators
The company also reports 44% more fraud detected using shared signals within the highest-risk 0.1% of sessions, with no increase in false positives. That is a result for a narrowly defined group, not a 44% improvement across every login. Nor does the release establish that this figure measures detection of malicious AI agents specifically. Those distinctions are essential when translating product-level performance claims into a bank’s business case.
Implementation will shape results. Feedzai’s partnership with IT Smart Systems shows the role of integration in bringing risk decisions into banking and payment workflows. Institutions adopting the expanded controls will need to calibrate them against their customers and channels, then compare fraud outcomes, unnecessary challenges and review workloads over a defined period rather than relying on a headline percentage.
The company’s work with Dock in Latin American payments likewise demonstrates that distribution can involve different operating environments. For agent detection, the useful evidence will be whether systems distinguish authorised from unauthorised activity without repeatedly blocking legitimate customers. Measuring false trust and excessive rejection together will provide a clearer assessment than counting how many automated sessions the system identifies.
Takeaway: Feedzai is extending bank-session risk assessment to AI agents while strengthening device-based authentication. The central challenge is distinguishing authorised delegation from malicious automation. Its reported fraud and authentication improvements offer potential business benefits, but their limited denominators and undisclosed testing details should not be treated as proof of agent-detection performance across all customers.
