Overview
Five major U.S. banks — Citibank, JPMorgan, Bank of America, PNC, and Citizens — have updated their terms of service to disclaim liability for any losses, unauthorised transactions, or account compromises arising from customers granting access to third-party AI agents. Critically, Citi’s updated policy extends this disclaimer to scenarios where the AI tool itself is hacked or leaks customer credentials — leaving users fully exposed with no path to reimbursement.
This marks a significant inflection point in the security posture of consumer-facing agentic AI. As AI agents proliferate in personal finance — capable of reading balances, executing payments, and managing investments — banks are drawing a hard line: the AI agent, and all associated risk, belongs to the customer.
Technical Analysis
The core security concern is credential delegation. When a user grants a third-party AI agent access to their bank account, they typically provide either direct login credentials or OAuth-style tokens. These credentials are stored within the agent’s configuration or memory layer, creating a persistent attack surface.
Threat vectors include:
- Credential harvesting from agent configuration: Attackers compromising the AI platform’s backend can extract stored banking tokens at scale (AML.T0083, AML.T0098).
- Tool poisoning: A malicious or compromised tool integration within an AI agent workflow could redirect transactions or exfiltrate account data (AML.T0110).
- Context poisoning: Injected instructions via documents or web content the agent processes could manipulate it into initiating fraudulent transactions (AML.T0080).
- Valid account abuse: Once credentials are obtained, attackers use them directly — no AI exploitation required (AML.T0012).
The absence of standardised, scoped, revocable permission models for AI agent access to financial systems is a foundational gap. Unlike Open Banking APIs with granular OAuth scopes, many AI agent integrations rely on full-credential sharing.
Framework Mapping
- LLM08 (Excessive Agency): AI agents with transaction execution capabilities represent an over-permissioned deployment pattern where the blast radius of compromise is disproportionate.
- LLM07 (Insecure Plugin Design): Third-party AI finance tools connecting to banking APIs without enforced read-only scopes exemplify unsafe plugin architecture.
- LLM06 (Sensitive Information Disclosure): Agent memory or configuration stores holding banking credentials are high-value targets for exfiltration.
- LLM05 (Supply Chain Vulnerabilities): The AI tool itself is a supply chain node — a compromised update or backend breach propagates directly to financial accounts.
Impact Assessment
The liability shift places the entire risk burden on individual consumers, who have little visibility into how third-party AI tools store, transmit, or protect their credentials. In a credential-compromise scenario, customers face financial losses with no regulatory backstop — existing consumer protection frameworks (e.g., Regulation E) were not designed with AI agent intermediaries in mind. The systemic risk scales with adoption: as millions of users connect AI tools to financial accounts, a single platform breach could yield mass account takeover.
Mitigation & Recommendations
- Revoke third-party AI agent access to bank accounts unless operationally essential; audit connected applications in your bank’s security settings.
- Use read-only integrations only — tools like Anthropic’s finance product that explicitly prohibit transaction execution reduce the blast radius of compromise.
- Prefer bank-native AI tools that maintain institutional liability coverage rather than third-party agents operating outside the bank’s terms.
- Demand scoped, revocable tokens: Advocate for and use platforms that support OAuth with minimal permission scopes rather than full credential sharing.
- Monitor account activity continuously and set low-threshold transaction alerts if any AI agent access is maintained.