Agent authentication

How a website or service checks that an artificial intelligence (AI) agent really is the agent it claims to be, such as a specific shopping assistant from a specific company. It is a separate check from confirming the person the agent works for.

A retailer might confirm that a request comes from a registered shopping agent before showing it checkout, much as a shop checks a delivery driver's company badge.

Proving which software sent a request with public-key cryptography, instead of trusting labels the client supplies, such as a User-Agent string or an Internet Protocol (IP) address. A typical check runs in four steps:

1. The agent's operator publishes a public key in a directory at a known web address.
2. The agent signs each request with its private key, adding a timestamp and a one-time value that stops replays.
3. The service fetches the public key and verifies the signature.
4. A valid signature lets the service apply its policy for that agent.

Several approaches are in use, often together:

• Signed web requests, such as Web Bot Auth, built on Hypertext Transfer Protocol (HTTP) Message Signatures (Request for Comments 9421) and still a draft at the Internet Engineering Task Force (IETF).
• Card-network agent programs, such as Visa's Trusted Agent Protocol and Mastercard Agent Pay, which register agents and have them sign requests showing whether they are browsing or buying.
• OAuth-based delegation, in which a person grants an agent limited, revocable account access without sharing a password.

Agent authentication sits next to terms that are easy to mix up:

• Agentic identity links an agent to a verified owner, sets its scope, and monitors it, while agent authentication answers only which agent is asking.
• Know Your Agent (KYA) verifies the agent, its mandate, and the person behind it, and agent authentication is its first check.
• Authentication confirms a person, while agent authentication confirms software.