Cloudflare Lists 14 Problems AI Agents Create for Websites. Sites Pay for Nearly All.

Cloudflare says agents that shop, log in and read for people leave websites holding the chargebacks, the serving costs and the bans, in a paper that also launches a coalition on agent identity.

A site that sees only the agent cannot prove who ordered a disputed purchase, and a won dispute still costs money. Serving an article to an agent brings no visit or ad view, and one abusive user behind a shared agent forces a ban that shuts out everyone else using it.

What are the 14 problems?

Friendly fraud, where a user approves an agent’s purchase and then disputes it, shows the pattern: a site may win such a dispute with the same device, IP address and delivery address as the user’s past undisputed purchases, and an agent’s purchase carries none of them.

Others hit identity and revenue. Identity checks assume the person present is the person verified. Agents that read articles or product data to answer a user elsewhere cost a site money to serve and bring no visit or ad view. When one user abuses a shared agent, the site can only block the agent, which shuts out its legitimate users too.

What would fix it?

The paper sets out four levels of user identity a site might require. The lowest is no identifier at all. Above it sit a session identifier, an account identifier that persists across sessions, and a unique per-user identifier from a third party. For the agent itself, Cloudflare prefers Web Bot Auth, which lets an agent sign its requests. Web Bot Auth cannot yet handle one agent working inside another, Cloudflare said.

Cloudflare also announced the Trusted Agent Delegation coalition, which will bring companies together to examine existing protocols and take problems to the Internet Engineering Task Force (IETF), which is already working on part of the problem. The paper names two example agents, Meta’s Muse and OpenAI’s dots, without listing either as a member.

Who has joined?

No one has been named. The paper lists no members, and neither does the coalition’s website. Cloudflare has set no meeting dates. Cloudflare says it has worked with major agent operators, payment providers, merchants and publishers, but does not say which have signed on.

What can sites do now?

Cloudflare customers can now see which agents visit and what each one does. They can allow an agent everywhere, block it only on pages that carry ads, or block it entirely.

Sites in the early-access Account Abuse Protection program can also see which agent made each login or sign-up. When one user abuses a shared agent, the site can write a firewall rule on the account’s Hashed User ID and leave the agent open to everyone else.

Cloudflare has invited companies that build agents and run cloud platforms to join, and none has been named.