PNC adds rules for third-party AI agents in online banking
Starting December 19, an AI agent can initiate a payment or transfer on a customer’s behalf only with PNC’s express permission for that service.
PNC Bank added rules for third-party AI agents to its Digital Services Agreement for online and mobile banking. The terms take effect December 19, 2026, and require agents to use connections the bank permits.
Customers can let an agent initiate, request, submit or complete a payment, transfer or other transaction only if PNC expressly allows it for that service. The AI-agent section names no service with that permission. Agents may help prepare a transaction where PNC allows it, and the bank may require customers to review, authenticate, approve or confirm it before processing.
Where PNC offers a connection, customers sign in directly with the bank and set which eligible accounts and information the agent can access, what it can do and for how long. PNC separately documents an account-linking interface that lets qualifying financial apps retrieve selected account data.
The agreement bars customers from sharing banking credentials, including one-time codes, with agents or letting agents request, collect, store or use them. Customers also cannot let agents use browser automation, screen scraping, or embedded or shared browser sessions to bypass PNC’s connection, security, access or transaction rules.
To the extent applicable law allows, the agreement makes customers responsible for access and activity they authorize or permit in breach of the restrictions, including unintended actions and actions beyond the authority they meant to grant.
Federal Reserve Governor Michael S. Barr said in a November 16 speech that major banks had reported attempts by AI agents to act in retail accounts in cases where the banks could not verify whether the agent was genuine, whether the customer had consented or whether the action matched the customer’s instructions.
The agreement allows PNC to block or delay agent-related access or instructions, or require additional authentication or confirmation, where applicable law and the relevant account or service terms permit it.
More news

Dragos publishes findings from about 2 million AI test prompts

OpenAI details GPT-6.2 safeguards against model extraction

OpenAI details frontier-model controls from training through deployment
