Cloudflare says websites should charge AI agents for access as Meta’s Muse learns to speak.

AI agents can search, shop and send messages for people. But when an agent visits a website, who pays for the work it creates?

Cloudflare says websites are carrying most of the cost. In a paper published on November 5, the internet security company argues that agents can cause problems for websites even when they are acting for real people.

An agent might buy something without making clear whether the person approved the purchase. It might try to log in to a bank using real account details. Or it might read a news story and share a summary without opening the page or seeing its ads.

“The Site bears the primary cost,” Cloudflare wrote in its paper, which lists 14 problems linked to AI agents. The company warns that agent traffic is growing quickly and could one day be greater than ordinary human browsing.

Cloudflare’s proposed rules

Cloudflare says websites should be able to understand who an agent is acting for and what it is allowed to do. It calls this idea the “Intermediation Principle”: a person’s relationship with a website, including security protections, should continue even when software acts between them.

The paper suggests several ways to help. Websites could use different levels of identity checks, from no identity information to a confirmed account. Cloudflare also supports Web Bot Auth to identify agents, and OAuth to let people give agents permission to use their accounts. It says websites need a new way to explain what agents may do, because the existing robots.txt system was not designed for this purpose.

Some Cloudflare tools are already available. Customers can see traffic from personal assistant agents and set rules to block certain agents on pages with ads. A tool to help protect accounts from abuse is in early access.

There are limits to the evidence. Cloudflare did not publish figures showing how many websites have these problems or how widely its tools are being used. The company also sells products for managing bots, so it has a business interest in this area.

Still, other findings point to questions about agent identity. Arkose Labs found that 97.7% of legitimate AI-agent sessions had no verified identity. Research firm HUMAN said Meta’s Muse made up 72% of AI-agent requests it observed.

Muse gains a voice and new tools

Meta has also been adding more features to Muse, its AI assistant. On November 5, the company began rolling out spoken conversations and animated avatars in the United States. The feature is not yet available in Canada. Users can choose from hundreds of voices, interrupt Muse, and let it continue working on tasks while they talk.

The next day, Meta introduced “spaces”: tools that users can set up by describing what they want in plain language. A space could track prices, collect saved Instagram items or help plan a wedding. Some examples are designed for shopping and selling, including a tool that brings together saved products, Marketplace listings and buyer messages.

Meta says spaces are private for now. It has not announced when users will be able to share them.

The new features could make Muse more useful in everyday life. But some details remain unclear, including how voice recordings are stored and whether information from spaces will be used for advertising. Meta has also not explained whether transactions made through Muse will include fees.

The question of responsibility

Cloudflare’s argument is that the growth of AI agents changes the relationship between websites and their visitors. A website may face more security risks or lose advertising income, while the person who asked the agent to act may not see those effects.

Meta’s expanding Muse shows why the issue matters. As agents take on more tasks, people, technology companies and websites will need clear rules about identity, permission and responsibility. For now, Cloudflare says, the costs often fall on the websites that agents visit.

What is Muse?

Muse is an AI assistant from Meta that people can interact with through conversation. Users can ask it questions, seek explanations, and get help with a range of everyday tasks, making it a useful way to explore information or work through a request. Its newer tools also connect the assistant with features across Meta’s services, including Instagram and Marketplace. These connections can help people use Muse in ways that relate more directly to the apps they already visit – for example, when they are browsing content or looking for items. The available features may depend on the service and how Muse is being used, but the goal is to make assistance more accessible within Meta’s platforms.