Citibank recently updated its service agreement, and if you are building in the intersection of fintech and AI, you need to pay attention. The bank has made it clear that if you use a third-party AI agent to manage your money and that agent goes rogue—or gets hacked—the bank isn't responsible. You are.
The Liability Shift
For years, the banking industry has operated on a relatively stable understanding of fraud protection. If someone steals your credit card number, there is a path to getting your money back. But as we move into the era of autonomous agents, the goalposts are moving. Citi’s updated terms essentially treat an AI agent like a physical key to your house. If you give that key to an AI bot and the bot decides to empty your account or gets tricked into doing so, the bank views that as an authorized transition.
This is a preemptive strike against the coming wave of autonomous finance. We are seeing a massive influx of startups promising to automate everything from bill pay to high-yield treasury management using LLMs. Citi is looking at this landscape and deciding they don't want to be the insurance policy for unproven software.
Why This Matters for Builders
If you’re a founder building an AI banking interface, this puts a massive target on your back. In the old world, you were a UI layer. In this new world, you are a custodial liability. If a customer connects your agent to their Citi account using their credentials, and your prompt injection protection fails, the customer loses their money and the bank points the finger at you.
This creates a chilling effect on innovation. For a builder, it means your insurance premiums and security audits just became the most important part of your tech stack. You cannot rely on the banking layer to act as a backstop for errors. If your agent misinterprets a "pay the balance" command as "send the balance to this random wallet," the transaction is considered valid because it originated from an authorized agent.
The Hack Loophole
Perhaps the most concerning part of the update is how it handles compromised credentials. Traditionally, if your login is leaked and a hacker uses it, you have protections. However, under these new terms, if those credentials are leaked through the third-party AI service you chose to use, the bank’s liability disappears. They are effectively saying that by choosing to use an AI intermediary, you are assuming all risks associated with that intermediary's security posture.
This is a classic legacy move. Banks love the idea of efficiency but they hate the idea of risk. By pushing the responsibility onto the consumer, they are forcing the market to self-regulate through fear. It forces the customer to vet the technical infrastructure of the AI tools they use—something 99% of customers are not qualified to do.
The Technical Debt of Trust
We have to look at the reality of how these agents work. They are often built on top of fragile APIs or, worse, through screen scraping. They are subject to hallucinations and adversarial attacks. When an agent acts on a bank account, it is doing so with the full permissions of the user. There is no "undo" button for a wire transfer sent by an AI that got confused by a phishing email.
For those of us in the crypto space, this sounds familiar. It’s the "not your keys, not your coins" mantra applied to legacy banking. But while crypto users generally understand that self-sovereignty implies self-responsibility, banking customers have been conditioned to expect a safety net. Citi is removing that net before the circus has even started.
The Opportunity in the Friction
Whenever a legacy giant creates a point of friction, there is an opportunity for builders. If the big banks won't vouch for AI agents, there is space for a new kind of financial institution—one built with agentic workflows in mind. We need protocols that can verify intent and set granular permissions for agents, rather than giving them a master key to the vault.
Instead of a binary "authorized" or "unauthorized," we need programmable money. We need the ability to tell a bank, "This AI agent is allowed to pay utility bills up to $500, but it cannot move money to a new recipient without a hardware-based 2FA check." Citi’s current approach is too blunt; it’s a total abdication of responsibility in the face of technological change.
A Warning to Founders
Do not ignore the fine print. If you are selling an AI financial tool, you are now selling a high-risk security product. Your marketing needs to be honest about where your liability ends and the bank's begins. If you tell users their money is safe and a Citi account gets drained because of a bug in your code, you aren't just facing a support ticket; you're facing a catastrophic legal event.
The era of "moving fast and breaking things" in fintech is over. When the bank explicitly tells you they won't cover the bill, you have to be the one to ensure nothing breaks. This means moving toward local execution of agents, encrypted credential handling, and perhaps most importantly, verifiable audit trails for every action an agent takes.
The Bottom Line
Citi’s move is a signal that the banking industry is terrified of the unpredictability of AI. They see the potential for systemic loss and they are closing the hatches. For the rest of us, it means the bar for building trust with users just got significantly higher. You can no longer hide behind the bank's fraud department.
The bank isn't going to protect your users from your code. If you're building in AI finance, security is no longer a feature—it is the entire product.
Read the original at Decrypt →