Cloudflare says agents break site-user trust
Cloudflare argues that remotely hosted AI agents hide the users behind them from websites, and says the web needs new identity, reputation and accountability mechanisms before sites can safely open up to agentic traffic.
Cloudflare has published a position paper arguing that personal AI agents take away websites' ability to recognize, serve and hold accountable the people they act for. The paper follows the company's recent work on agent traffic, which we covered when it added per-agent controls.
The post covers agents hosted by an "Operator" that control a cloud computer or browser for a user. It names Meta's Muse and OpenAI's dots as examples. Local agents such as OpenClaw, and agents running in the user's own browser, are explicitly out of scope. Cloudflare treats hosted agents as the newest type of intermediary, alongside cloud browsers and proxies. It says agentic traffic "may eclipse traditional human browsing" soon. If that happens, sites can no longer just block intermediaries.
At the center of the argument is what Cloudflare calls the Intermediation Principle: the business relationship between users and sites, and each side's security responsibilities, must hold whether or not an intermediary sits between them. The company says that means three things: identity, reputation and accountability, for both the user and the agent's operator.
The post lists problems that were once solved but come back when agents act for users:
- Account sharing: one paid subscription can serve many users.
- Compliance: agents don't give sites enough assurance for age or geographic checks.
- Friendly fraud: a user can dispute a purchase their agent made. The site then lacks the usual device and IP evidence it would use to show the user authorized it.
The source text available to us is cut off partway through. It does not show what technical mechanism, if any, Cloudflare proposes.