Oleksandr Kuzmenko, EditorPublished 6 min readAI-assisted · editor-reviewed
TL;DR
- US banks have officially shifted 100% liability for AI agent errors and breaches onto end users
- Third-party agents that leak credentials or trigger accidental transactions will not be protected by bank fraud guarantees
- Engineers building financial AI tools should strictly follow the read-only pattern pioneered by Anthropic
Key facts
- Impacted Institutions
- Citibank, JPMorgan, Bank of America, PNC, Citizens
- Policy Scope
- Customer liable even if AI tool leaks credentials or is hacked
- Anthropic Model Precedent
- Read-only access (cannot move, manage, or invest money)
Banking Terms Shift Risk to End Users
Citibank, the third-largest commercial bank in the United States, updated its Online User Agreement to establish that customers who share banking credentials or authorize third-party AI tools bear sole responsibility for all resulting activity. Crucially, the policy dictates that if an external AI service leaks credentials or gets compromised, the customer remains on the hook for unauthorized transactions. This follows similar policy revisions from JPMorgan, Bank of America, PNC, and Citizens.
The Read-Only Pattern for Financial Agents
Financial institutions view credential-sharing agents as both liability hazards and competitive threats to deposit margins. For developers building agentic workflows, the institutional backlash confirms that requesting write or transaction privileges is an anti-pattern. Anthropic avoided this trap by designing its native personal finance integration to be strictly read-only, specifying that the model cannot move, invest, or transfer funds.
Design Rules for Model Context Protocol (MCP) Tools
Engineers developing autonomous tools for finance or sensitive third-party services should enforce architectural guardrails:
- Restrict all agent tools to read-only scopes (e.g., balance and transaction history queries).
- Never store or transmit raw banking credentials; rely on regulated aggregation tokens where available.
- Require out-of-band, multi-factor human confirmation for any downstream state-modifying action.
Try it in 2 minutes
{"name": "get_account_balance", "description": "Fetch account balances in read-only mode", "parameters": {"type": "object", "properties": {"account_id": {"type": "string"}}, "required": ["account_id"]}}When to use
- Designing agent tool definitions that query account balances and categorized spend
- Architecting Model Context Protocol (MCP) connectors for personal finance and accounting workflows
When NOT to use
- Autonomous transactional agents with write permissions to initiate bank transfers or bill payments
- Sharing raw client online banking passwords or multi-factor authentication sessions with LLMs
What to do today
- Audit all agent tools to ensure strict read-only query permissions on financial endpoints
- Remove any automated transaction execution capabilities from consumer-facing agent workflows
- Implement out-of-band manual confirmation if your agent prepares payment orders
Step names are taken from What to do today.
Step names are taken from What to do today. Audit all agent tools to ensure strict read-only query permissions on financial endpoints. Remove any automated transaction execution capabilities from consumer-facing agent workflows. Implement out-of-band manual confirmation if your agent prepares payment orders
- Audit all agent tools to ensure strict read-only query permissions on financial endpoints
- Remove any automated transaction execution capabilities from consumer-facing agent workflows
- Implement out-of-band manual confirmation if your agent prepares payment orders
