Dev News Daily ENDE

Cloudflare argues sites need to see the user behind an AI agent, not just the agent

Cloudflare published a long post on 5 November about how websites should treat personal AI agents, the hosted services that take instructions from a person and act in a browser or cloud computer on their behalf. Its starting point is that a remotely hosted agent is an intermediary: the site sees the agent, not the user, and one agent may act for many users at once.

The company proposes what it calls the intermediation principle: the business relationship between users and sites, and each side's security responsibilities, must be upheld whether or not an intermediary sits in between. In security and fraud prevention terms, that means a site still needs identity, reputation and accountability, and must be able to answer whether the person acting is who they claim to be, whether they can be trusted and who answers for the action.

The post walks through what breaks today. One paid subscription can serve many users through one agent. Friendly fraud, where a user approves a purchase and later disputes it, gets harder to fight because the usual evidence of the same device, network and address is missing. When one user abuses a shared agent for credential stuffing or card testing, a site can only block the agent, which also blocks its legitimate users. Content sites pay to serve pages that an agent reads and summarises elsewhere, with no visit, ad view or purchase in return. In nearly every case, Cloudflare writes, the site bears the cost.

It also rejects two common fixes. Passing a CAPTCHA through to the user proves only that some human was present at that moment and adds friction that operators will try to remove. Narrow permission scopes tell a site little, because agents usually ask for permission to do everything, and a scope says nothing about who is responsible when the agent gets something wrong.

Instead, the post calls for protocols that let sites and agents agree on what is allowed. It cites MOLE, an upcoming approach to rate limiting without user-specific identifiers, and the x402 protocol for privacy-preserving payments on the web. Agents running locally on the user's own device are explicitly outside the scope of the argument.

Cloudflare argues sites need to see the user behind an AI agent, not just the agent
Cloudflare argues sites need to see the user behind an AI agent, not just the agent — Dev News Daily

What it means

Cloudflare sits between many sites and their traffic, so its framing is likely to shape how agent traffic is classified and charged. Sites can expect pressure to move from blocking agents to accepting signed, accountable agent traffic.

Written by Victoria Shinder.