AI agents now book, buy and research for millions of people. To a website, though, an agent is a stranger: it arrives from a data centre, it may act for many people at once, and it does not get tired. On 5 November 2026, Cloudflare, which serves a third of the top 10,000 websites, published its answer: agents should say who they are and who they act for. For teams building AI agents or running automation of any kind, it shows where access rules are heading.
What Cloudflare announced
The post, “Agents aren’t humans (but they kinda are)”, does three things:
- Sets out a principle. “The business relationship between Users and Sites, and each participant’s security responsibilities, must be upheld regardless of intermediation.”
- Launches a coalition. Trusted Agent Delegation (TAD) gathers “major Operators, identity verification companies, payment providers, merchants, publishers, proxy providers, and cloud browsing companies” to define the problems and take them to the IETF.
- Ships three features for Cloudflare customers, covered below.
Cloudflare calls agents intermediaries, in the same family as cloud browsers and proxies. What has changed, it argues, is scale: “Agentic traffic is growing rapidly and may eclipse traditional human browsing on the Internet in the near future.” Once that happens, simply blocking intermediaries stops being an option.
The problem: sites see the agent, not the user
Cloudflare lists more than a dozen things that become harder when an agent sits in the middle. A selection:
| Challenge | What goes wrong |
|---|---|
| Account sharing | One paid subscription serves many users, and the site sees one client |
| Friendly fraud | A user approves an agent’s purchase, then disputes it, and the usual evidence of device and IP does not exist |
| Traffic commingling | One user abuses a shared agent, and blocking the agent blocks all its legitimate users |
| Resource exhaustion | Agents poll stock and prices around the clock and use up free trials |
| Uncompensated consumption | The agent reads articles, product data or reviews and answers elsewhere, so the site gets no visit, ad view or sale |
| Identity spoofing | Agents run on residential IPs and real browsers, and attackers copy a known agent’s user agent |
The post also dismisses two common fixes. Passing a CAPTCHA to the user proves a human is present at that moment “and nothing more”. Narrow permission scopes do not work either, because agents ask for permission to do everything.
Four levels of identity
Cloudflare’s answer is a sliding scale of how much an agent discloses about its user, chosen by the site according to risk:
| Level | What the site learns | Example use |
|---|---|---|
| Nothing | No identifier | Reading a public page, like private browsing |
| Session | That requests belong to one session | Building reputation within a visit |
| Account | A stable identifier for the same user across sessions | Linking a site account to an agent account |
| Unique | A per-person identifier from a third party, across agents and operators | High-risk actions or strict compliance |
The post stresses data minimisation: in many cases a site only needs to know whether to trust a request, not who sent it. For that, it points to privacy-preserving tools such as Privacy Pass and MOLE, a proposed successor to Private Access Tokens that aims to allow rate limiting without user identifiers.
The building blocks
Most of the pieces already exist or are in progress:
- Web Bot Auth, Cloudflare’s “preferred method” for agent identity. The agent signs each request with HTTP Message Signatures (RFC 9421), so the site can verify which agent, and implicitly which operator, sent it. Cloudflare introduced it in May 2025 and created a signed agents category in August 2025. Our post on AI agents and bot walls covers how signing works today.
- Claims. Some actions need a fact about the person, such as being over 18 or being the cardholder, which the operator or a third party may already have verified.
- Payments, through protocols such as x402, which aims to allow privacy-preserving payments on the web.
- Access policies. Cloudflare notes that “Robots.txt says where a crawler may go, but isn’t fully suited to express what Agents may do.” The IETF’s AI Preferences working group covers preferences for uses such as training and search, but no group yet covers what agents may do on a site.
What Cloudflare customers can do now
As of the week of 5 November, Cloudflare customers can:
- See which agents visit. Security Analytics now shows personal assistants as their own traffic category, with each agent’s requests, paths and errors.
- Set access per agent. A site can allow an agent everywhere, block it only on pages with ads, or block it entirely, separately from its search and training settings.
- Act on the user, not the agent. For sites in the Account Abuse Protection early access, every login and sign-up shows whether a personal agent made it and which one. A site can then act on one account without blocking the agent for everyone else.
The second point matters most for anyone running automation. Access is becoming a per-agent, per-page decision, not a single yes or no.
What it means for teams that use automation
- Agents and data collection are different things. Cloudflare’s proposal is about agents acting for a person, often logged in and buying things. Collecting public pages such as prices, search results or ads is a different job with different rules, but site owners will increasingly use the same tools to tell them apart.
- If you run an agent product, plan to sign. Verifiable identity is the route Cloudflare favours to reliable access, and agents that do not identify themselves will be easier to group and restrict.
- Expect page-level rules. A site may welcome an agent on product pages and block it where ads pay the bills. Build agents that handle a refusal gracefully instead of retrying.
- Blending in is the problem being solved. Cloudflare names agents that run on residential IPs and real browsers and avoid standing out as a reason sites lose control. A residential IP shows a page as a local visitor sees it; it is not a way to act as someone else or to get around a site’s rules. Our guide to ethical residential proxies for AI data collection covers where we draw that line, and our AI shopping agents post covers what changes for product data.
What is still open
Much of this is a proposal. TAD is a coalition, not a standard, and Cloudflare says it will hold side meetings at the IETF and workshops later. Agents running on a user’s own device are outside its initial focus. Privacy questions remain too: an identity scale only protects users if sites ask for the lowest level they need. Still, with one company serving so much of the web, its customer features tend to become the practical rules long before any standard is finished.
The bottom line
Cloudflare wants AI agents to stop looking like anonymous visitors and start saying who they are and who they act for, with sites choosing how much to ask. The standards are still being written, but per-agent controls are already live for Cloudflare customers. Teams building agents should plan for signed requests and page-level rules now.
Sources and references
- Cloudflare, Agents aren’t humans (but they kinda are) (5 November 2026): blog.cloudflare.com
- Cloudflare, Forget IPs: using cryptography to verify bot and agent traffic (15 May 2025): blog.cloudflare.com
- Cloudflare, The age of agents: cryptographically recognizing agent traffic (28 August 2025): blog.cloudflare.com
- IETF, RFC 9421, HTTP Message Signatures: rfc-editor.org
- IETF, AI Preferences working group: datatracker.ietf.org