Know Your Agent

Know Your Agent (KYA) is the process of verifying an AI agent’s identity, the person or business accountable for it, and the actions it is authorized to take before it can act on your platform. 

A customer can now use an AI shopping assistant to purchase products. The merchant now needs to know whether the purchase is made by software on behalf of a real customer or whether it even has the permission to complete that purchase. 

KYA answers three questions about that software: whether the agent is authorized, which verified person or business is behind it, and whether the action the agent is performing falls within the permissions it was given. It does not replace KYC or KYB since the principal behind it is accountable and still has to be verified the same way any customer would be.

KYA verification process: 

KYC tells about who the customer is, not about the agent. It doesn’t tell whether the agent acting in their name belongs to them or what it may do. 

KYA actually closes that gap: 

• Agent Identity: a unique credential that is bound to the agent
• Principal (Owner): a verified person or business legally accountable for the agent 
• Delegated Authority: checks the permissions that are given to the agent
• Audit Trail: a log that links every action back to the agent and its principal

A common failure that exists is treating the agent as the customer. An agent that has a valid token but with no verified principal is considered an anonymous account. 

Difference between Non-human identity, machine identity, and agentic identity: 

Each term has a different meaning and covers a different layer of risk.

Machine identity confirms the software is genuine. Only agentic identity confirms who is accountable for it. 

AI agent fraud and KYC:

AI agent fraud happens because traditional KYC was built for verifying people (customers), not autonomous software agents. Agents act at machine speed with whatever credentials they hold, so fraud often looks legitimate. 

Common examples of AI Agent Fraud: 

Stolen Agent Credentials: Someone steals an AI agent’s access key or login. They would act like moving money or accessing accounts. This activity may look normal, but it comes from a valid credential. 

AI Agent opening fake accounts: Fraudsters use the automated accounts to create different accounts using fake or stolen identities. The volume and speed make the detection harder. 

Prompt Injection Attacks: Fraudsters usually manipulate an AI agent with hidden instructions. The agent would be trained to perform actions it was not supposed to do. For instance, an AI assistant approved a refund though rules did not allow refunds above a certain amount. 

Fake agents pretending to be trusted agents: A malicious agent imitates a legitimate one. It attempts to convince systems to approve payments or access sensitive information. 

Related rules and standards:

KYA is a practical control framework, not a requirement named in the rules. The existing rules may affect the AI disclosure, payment authentication, automated decisions or risk assessment, depending on the services, jurisdiction, and the company’s role. 

A practical KYA check runs in five steps: 

1. Register the agent and issue it a unique, revocable credential
2. Verify the principal through KYC or KYB checks that are applicable to the service and risk. 
3. Link the agent to the principal and record a mandate specifying permitted actions, limits, expiry, and who may revoke it.
4. Before each requested action, check the agent’s credentials and the current mandate, including its limits and revocation status. Seek fresh customer approval when the action falls outside the mandate or meets a defined risk threshold.
5. Record the decision and monitor subsequent activity for misuse.

KYA check at checkout allows or blocks the payment: 

When the shopping agent is trying to pay, the merchant first checks the agent’s credentials, then links them to the verified customer who authorized it. It then compares the purchase against that customer’s mandate. These are approved by the merchant: the spending limit and when the permission given to the AI agent will expire. 

When Allowed: The agent buys an item that is approved by the merchant within the customer’s $100 limit, before the mandate expires. The payment goes ahead, and the decision is logged. 

When Declined: On the other hand, if the agent tries to spend $250, or pay a merchant the customer never approved. The payment is either blocked or sent to the customer for approval. 

The customer can revoke the agent’s access at any time, and every decision is recorded against the agent and the customer.