What Cloudflare published

On 5 November 2026 Cloudflare posted a whitepaper called Agents aren't humans (but they kinda are) and launched a coalition called Trusted Agent Delegation (TAD). The coalition wants open standards that help websites recognise AI agents acting for people. The problem the paper describes is familiar to anyone who has built a browsing or shopping agent. A website sees the agent's traffic and can't see the person behind it, so the site loses the identity, reputation and accountability signals that its fraud and security controls depend on. The paper lists problems such as account sharing, friendly fraud, identity spoofing and traffic commingling, which is when many users' requests arrive mixed together through one agent. The paper concludes that the site bears most of the cost in nearly every case. Its one-line summary of the risk is blunt: "Agents are able to scale human access almost without limit."

Cloudflare's answer is what the paper calls the Intermediation Principle. The relationship between a user and a site, along with each party's security responsibilities, should stay the same when an agent sits in the middle. Biometric Update's report on the launch describes TAD as resting on identity, reputation and accountability, with Tools for Humanity (the company behind World ID) as a launch partner.

Graduated disclosure

The core idea is that an agent should reveal only as much about its user as a given action needs. The whitepaper defines four tiers:

  • Nothing: the site judges the agent alone, with no user identifier.
  • Session: an identifier links requests within one session to a single agent instance, so a site can act against that session without blocking the agent's other users.
  • Account: a stable identifier from the operator or a third party links one user's requests across sessions.
  • Unique: a persistent per-user identifier that holds across operators and agents, such as a government ID or a proof-of-personhood credential, reserved for high-risk actions.

Accountability works through the operator. According to Biometric Update, when a site reports abusive users of an agent, the operator is expected to deal with those users so the site doesn't have to block the whole agent. The same report says Cloudflare will ask for evidence of these controls before it grants verified agent or operator status.

Web Bot Auth does the signing

TAD doesn't define a new signing scheme. The whitepaper names Web Bot Auth (WBA) as the preferred way for an agent to identify itself and, by extension, its operator. WBA uses HTTP message signatures, which means the agent signs selected parts of each HTTP request with a private key and the receiving site checks the signature against a published public key. Cloudflare's Web Bot Auth documentation sets out the mechanics. Cloudflare accepts Ed25519 keys only. The agent's operator hosts a JSON Web Key Set at /.well-known/http-message-signatures-directory over HTTPS, and that directory response must be signed too, so nobody else can copy the directory and register on the operator's behalf.

The docs show key generation with OpenSSL, followed by a conversion to JWK format with the jwker tool:

openssl genpkey -algorithm ed25519 -out private-key.pem
openssl pkey -in private-key.pem -pubout -out public-key.pem
jwker public-key.pem public-key.jwk

Each signed request then carries a Signature-Agent header pointing at the key directory, a Signature-Input header listing the signed components with tag="web-bot-auth", and the Signature header itself. The documentation flags several traps. The expires value should be short, around a minute. Cloudflare doesn't track nonces, so the short expiry is the main defence against replayed requests. Some signature components and parameters, including @query-params, cause verification to fail outright. Registration happens through the Bot Submission Form in the Cloudflare dashboard, with "Request Signature" selected as the verification method.

Privacy through MOLE

The obvious worry with agent identity is cross-site tracking. The whitepaper points to privacy-preserving tools such as Privacy Pass and an upcoming protocol called MOLE, which it describes as aiming for rate limiting without user-specific identifiers. The MOLE architecture draft, dated 3 November 2026, has authors from Google, Mozilla and Cloudflare. In MOLE, a third party called an Anchor gives a client a limited number of endorsements, perhaps after a login or a CAPTCHA. The client redeems an endorsement once at a site's Moderator in exchange for a credential. The Moderator can then raise, lower or rate-limit that credential's access without learning who the client is or which Anchor vouched for the client. The draft is an Internet-Draft with informational status, and the page doesn't say whether an IETF working group has adopted it.

What's live and what's still a proposal

Some Cloudflare features shipped alongside the paper. According to the whitepaper, Security Analytics now shows Personal Assistants as a separate traffic category, with a widget listing each agent, its request count, the paths the agent hits and the errors the agent runs into. Site owners can set access policies per agent, including blocking an agent only on pages that carry ads. Account Abuse Protection, in early access, shows whether a personal agent made a login or sign-up request. Sites can also write WAF rules against a Hashed User ID, which lets them stop one abusive account using a shared agent without cutting off everyone else on that agent.

TAD itself is a proposal. The coalition site offers an email list and doesn't name members or give dates. The whitepaper admits that chains of intermediaries, such as one agent calling another inside a cloud browser, still have no defined way to identify themselves. Biometric Update reports that Tools for Humanity plans to update its agent delegation SDK "in the coming weeks and months", and that Cloudflare intends to tighten verified agent requirements without giving a timeline. Neither source gives pricing for Account Abuse Protection or for the new analytics features.

What builders here can do now

New Zealand teams building agents that browse, book or buy on behalf of users will meet Cloudflare-fronted sites whether or not those sites are local. Signing requests with Web Bot Auth costs little more than a key pair and a hosted JSON file, and the registration form is already open. Teams that adopt WBA now will be ready for whatever verified-agent requirements arrive. Those teams should also build session identifiers into their agent design early, because the Session tier asks the operator to tie requests to one agent instance. Operators that can't suspend a single user will struggle to meet the accountability expectations the coalition describes.

Teams running sites behind Cloudflare can check the Personal Assistants category in Security Analytics to see how much agent traffic they already get before deciding on per-agent policies. Anyone with views on how disclosure tiers should work for NZ services, where a government ID check is a heavy requirement for a small purchase, can join the coalition's email list and respond while the framework is still open for feedback. If you build a signed agent or a site policy around WBA, we'd like to see the result on our projects page.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member