Customers these days can log in to their bank account through a mobile app or website, making online banking the preferred method for convenient access. But behind these digital applications are application programming interfaces that malicious actors can target, potentially with more precision with the help of agentic artificial intelligence.
Hidden APIs can be a risk for financial institutions, similar to shadow IT and other parts of the environment that have reduced visibility. A Barracuda Networks blog calls these unclaimed elements “zombie APIs” — no longer in active use, or even remembered, but still providing a point of access to an organization’s system.
“Zombie APIs are dangerous because they often have the weakest ownership and visibility,” says Adam Khan, vice president of global SecOps and AI security for extended detection and response/managed detection and response at Barracuda Networks. “A forgotten API no longer has to be easy to exploit. It just has to stay exposed long enough.”
Financial organizations must improve API inventory and visibility and manage control of their authorized AI agents so that any unauthorized attempts don’t snowball into something bigger.
Click the banner below to find security and compliance support for financial institutions.
Inventory and Visibility Are Key Components for Securing APIs
While it might seem easy to discuss APIs in terms of averages, David Giambruno, vice president for AI and business transformation at Forcepoint, says that the number of APIs an organization relies on will differ and that there shouldn’t be an invented industry average.
“What you actually find is the gap between the environment people think they run and the one that exists,” Giambruno says. “Reconcile three planes: what is declared in specs and catalogs; what is deployed in clusters and partner tenancies, including test and stage that still route; and what is observed at runtime, including batch jobs and agents.”
API inventory can be kept as simple as answering these questions:
- What can the API reach?
- Who owns it?
- What depends on it?
- Which identities can call it?
- What data can it expose?
- What business actions can it take?
Khan notes that at a large financial institution, discovery can uncover dozens of unknown or poorly tracked endpoints, especially if there has been a merger or acquisition and legacy systems are involved.
“The number isn't the real problem. The real risk is the API nobody can confidently tell you who owns, what data it touches or whether anyone is monitoring it,” Khan says.
Sensitive account changes and recovery flows are classic API authorization and authentication failure modes, Giambruno adds, highlighting a hypothetical scenario in which an older API allows an attacker with a compromised session to change recovery information without a fresh challenge. An agentic AI-powered account takeover via API doesn’t have to look like someone cracking the mobile app; it can look like the back end accepting a request it should have refused.
Click the banner below to discover how to strengthen cyber resilience in finance.
Endpoints that can’t be deprecated should be considered “high-risk legacy assets,” Khan says, with compensating controls such as strong authentication, least privilege, tightly restricted access, rate and behavioral controls, full logging and a clearly assigned owner. This can help minimize the damage should a breach attempt occur.
Authorized AI Agents Require Guardrails
Giambruno advises banks to scrutinize their own AI agents as much as they anticipate rogue agents from a malicious actor.
“An authorized agent can still be overpermissioned. It can see more than the task requires and take actions that should demand another approval,” he says. Organizations should track ownership and related actions, narrowly delegate permission and limit data access. Platforms such as Forcepoint AI Data Security can help with attribution, policy enforcement and control over what agents can reach in connected apps.
It’s not that every attack becomes an AI attack, Khan says, but that discovery, testing and persistence now happens continuously and at machine speed. “Defenders can't manage a real-time attack surface with a periodic inventory. AI isn't rewriting the fundamentals of security; it's shortening the time we have to execute them,” he adds.
