IA · 18 November 2026 · 5 min read

If an AI Agent Makes a Mistake or Gets Hacked, You Pay: Major US Banks Shift Full Liability to Users

In brief: Major US commercial banks, including Citibank, JPMorgan Chase, Bank of America, and PNC, have revised their online agreements to hold customers entirely liable for any transactions executed by third-party AI agents. Under the updated terms, users lose standard fraud reimbursement even if the external AI platform suffers a data breach or leaks their banking credentials. The move highlights a sharp institutional backlash against the deployment of unvetted autonomous agents across financial networks.

by Team Mocchi's

If an AI Agent Makes a Mistake or Gets Hacked, You Pay: Major US Banks Shift Full Liability to Users

Erasing Decades of Fraud Protection

For decades, commercial banking has relied on a foundational understanding: when unauthorized transactions, security compromises, or fraud occur without gross negligence by the customer, regulatory frameworks and banking safeguards step in to reimburse the lost funds. The rapid deployment of autonomous artificial intelligence agents into everyday finances has suddenly unraveled that presumption.

This week, Citibank—the third largest commercial bank in the United States—formally updated its Online User Agreement. The revision explicitly informs customers that anyone who grants bank credentials or portal access to third-party AI financial tools assumes sole economic and legal liability for the outcome. If an agent executes an erroneous wire transfer, makes unintended investment decisions, or falls victim to automated fraud, the bank will not reverse the charges or cover the loss.

Wall Street Forms a United Front

Citibank's contractual revision is not an isolated policy, but part of a coordinated defensive stance adopted across Wall Street. As reported by TechCrunch, identical restrictions were implemented last month by the country’s two largest lenders, JPMorgan Chase and Bank of America. Joining them today was PNC, the eighth largest US bank, alongside regional powerhouse Citizens Bank, which issued formal email warnings to account holders.

A Citi spokesperson told reporters that the policy update was designed to bring clarity to the inherent risks of sharing banking credentials with third-party software. From the perspective of financial risk officers, an autonomous agent navigating an online banking interface is legally indistinguishable from a user handing out an open power of attorney. Consequently, every command executed by an agent is treated as fully authorized by the customer, stripping away the safeguards established under federal electronic fund transfer regulations.

The Breach Clause: Liable Even When the AI Is Hacked

The most aggressive provision in the updated agreements concerns external cybersecurity failures. Citibank’s terms specify that user liability extends well beyond prompt errors or algorithmic hallucinations: if the third-party AI vendor suffers a cyberattack, leaks session tokens, or exposes stored banking passwords, the customer remains entirely liable for any fraudulent activity conducted by hackers on their account.

This approach marks a dramatic departure from standard consumer protection principles in digital finance. Typically, breaches occurring at accredited third-party intermediaries do not leave the end consumer personally liable for systemic compromises. By enforcing this clause, financial institutions are shifting the systemic technological risks of generative AI models and startup ecosystems directly onto the end user.

Agentic Ambitions Clash With Banking Infrastructure

The banking sector's defensive maneuver comes as frontier AI labs and fintech developers push aggressively toward autonomous financial delegation. Tools designed to monitor spending, negotiate bills, optimize corporate cash flows, and initiate scheduled payments have been marketed as the next frontier of personal and business productivity.

Yet many of these services continue to bypass formal open banking APIs and granular permission controls, relying instead on screen-scraping techniques and stored user credentials to interact with legacy banking web interfaces. For banking security teams, non-deterministic agents vulnerable to prompt injections and credential harvesting represent an intolerable threat vector to national clearing systems.

Mocchi's take

For European and Italian companies operating under PSD2 and the upcoming PSD3 regulations—where credential sharing and screen-scraping are already restricted in favor of Strong Customer Authentication (SCA)—the US banking reaction provides a vital lesson in risk governance. Enterprises building custom software or deploying automated workflows must recognize that delegating financial execution to autonomous agents without strict human-in-the-loop oversight invalidates core liability protections. In our view, the proper role for AI agents in financial operations is reconciliation, analysis, and transaction staging; final execution and authorization must always remain bound to human cryptographic validation under strict least-privilege principles.

Further reading

All articles on the Mocchi's blog