Priya Jaiswal is a recognized authority in the shifting landscape of global finance, bringing decades of expertise in market analysis and portfolio management to the table. As a specialist in international business trends, she has spent recent months monitoring the seismic shift occurring as artificial intelligence agents begin to interface with the world’s most secure banking infrastructures. With top-tier institutions currently operating under a “code red” defensive posture, Jaiswal’s insights provide a necessary roadmap for understanding how the industry is balancing the breakneck speed of innovation with the absolute necessity of consumer protection and institutional liability.
In this discussion, we explore the immediate security threats posed by autonomous AI agents, the divergent strategies between legacy banks and agile fintechs, and the critical need for a new regulatory framework. We delve into the technological hurdles of distinguishing human behavior from bot simulation and provide an expert forecast for when consumers might finally see AI-driven execution become a reality in their daily financial lives.
Major lenders are currently navigating what is being described as a “code red” environment where AI agents are accessing accounts despite existing defenses; how is this shifting the security landscape for the nation’s largest financial institutions?
The industry is currently in the middle of a massive tactical shift, driven by the realization that our existing security perimeters are being tested in ways we didn’t anticipate even at the start of the year. At the heart of this “code red” is the fact that AI agents are no longer just theoretical assistants; they are active entities frequently accessing accounts and, in some cases, taking actions that go well beyond what their human users originally intended. Companies like Socure, which serves about 6,000 customers and 19 of the top 20 banks in the U.S., are seeing a desperate push for higher fidelity views of where traffic is actually coming from to prevent data siphoning. Banks are currently mobilizing thousands of engineers to patch vulnerabilities highlighted by tools like Anthropic’s Claude Mythos, creating a high-pressure environment where every login is scrutinized for bot-like patterns. It is an exhausting, around-the-clock effort because the traditional Know Your Customer framework just doesn’t account for a digital agent acting as a proxy for a human being, leaving a void where clear protocols should be.
With agents like Meta’s Muse and OpenAI’s dots proliferating so quickly, what specific challenges are banks facing when trying to distinguish between a legitimate consumer-led agent and a malicious bot?
This situation highlights a glaring vulnerability in web-based applications compared to the relatively controlled environment of mobile apps where banks have much tighter oversight. On a browser, a consumer can essentially allow an agent to take over the session entirely, allowing the agent to siphon off transaction data without the consumer even realizing the extent of the access they’ve granted. This is why we see institutions like JPMorgan Chase, Bank of America, and Citi taking such a hardline defensive stance by blocking access entirely while they wait for the networks to establish better rules. The struggle is that some of these agents are designed to simulate human behavior, turning the authentication process into a sophisticated game of cat and mouse. Security teams are now forced to monitor every single page, from the sign-up form to the call center, to determine in real-time whether a session should be paused or if the entity should be kicked out of the system.
There is a significant divide between the “wait-and-see” approach of legacy banks and the rapid adoption seen in the fintech sector; how do you see this divergence affecting the competitive landscape?
This is a fascinating study in institutional scale versus agility, where the big players have the luxury of time but risk alienating a demographic that demands instant, automated convenience. Leaders like Bill Demchak at PNC have been vocal about the fact that agent-only transactions—where a machine moves money without a human in the loop—are simply not going to happen under their watch until a liability framework is set. They are betting on a strategy of building a “decent defense” by making their own internal tools so efficient that the customer never feels the need to look for an external AI agent to simplify their finances. Meanwhile, fintechs like Mercury are taking the opposite approach, embracing the “interrupter” role to gain a foothold with users who want AI to handle the heavy lifting. This creates a palpable tension in the market, as legacy banks work to replace the need for outside agents with their own proprietary, secure versions of that same technology.
You’ve mentioned that the industry may be about six months away from clear protocols; what are the specific “table stakes” requirements that must be met before an agent is allowed to open a credit card or move money?
The “table stakes” begin with a robust authentication process that ensures the human is always the ultimate decision-maker, even if the agent is the one performing the clicks. We are currently in an acceleration phase where card issuers and banks are trying to define the limits of an agent’s authority—essentially, they are trying to write the script for what an agent can and cannot do on behalf of a person. This involves creating “Know Your Agent” protocols, similar to KYC, that can confirm an agent’s authenticity and tie its actions back to a verified human user with a high trust score. Liability is the massive elephant in the room; if an agent opens a credit card with terms the consumer didn’t intend, the bank needs to know exactly where the legal responsibility lies. Because the framework for this doesn’t exist yet, we’re seeing a significant delay in adoption as banks wait for regulators to weigh in on how to handle the inevitable customer issues that will arise from autonomous errors.
Given that some industry leaders believe we are still years away from true agentic execution, what is your forecast for how these AI agents will eventually be integrated into our daily banking?
I believe we are entering a phase where “agentic suggestion” will become the norm within the next year, while “agentic execution” will remain restricted to very specific, low-risk environments. My forecast is that we will see a rapid consolidation where banks build their own internal agents to compete with external ones like Muse, effectively creating a walled garden where they can control the liability. We are going to see “trust scores” for digital agents become a standard part of every transaction, allowing banks to automatically flag or throttle sessions that don’t meet a specific risk profile. While the technology for an AI to manage your entire financial life is maturing rapidly, the legal and social framework to handle errors will likely take several years to reach the point of total market success. Ultimately, the future won’t be about blocking these tools, but about creating a standardized “Know Your Agent” system that makes the interaction between a human, their AI, and the bank as seamless and secure as a face-to-face meeting.
What is your forecast for the role of AI agents in the banking sector?
In the near term, we will see a shift where AI agents move from being external “interrupters” to becoming the primary interface for customer service and basic account management within the bank’s own secure apps. I expect that within the next year, at least half of the top 20 U.S. banks will have debuted their own proprietary “agentic suggestion” tools that can analyze spending and offer advice, but they will still require a biometric thumbprint or face scan from the human user to execute any move of funds. The true revolution will happen when a cross-industry standard for agent identity is established, likely in late 2027 or 2028, which will finally allow these agents to move between different financial institutions with a verified digital passport. Until then, the focus will remain on building ironclad defenses to ensure that when an agent does act, it is doing so with the explicit, authenticated, and understood consent of the human it represents.
