Security · AI agents

AI agents are already visiting your website. Start counting them

Customers are beginning to send AI agents to browse, book and buy for them. Most Malaysian websites cannot tell those agents from people or from abuse. How to log and classify agent traffic now, on any stack, and set a policy for each agent.

Pikabyte Engineering 7 min read

4 checks

identify AI agent traffic on the web server, gateway or firewall you already run

Your customers are starting to send AI agents to do things for them: compare prices, book a table, fill a cart, check a balance. When that agent reaches your website, you see the agent. You do not see the person behind it.

Cloudflare set out the problem in a post on 5 November 2026. One agent acts for many users, so the site cannot identify, serve or hold accountable the person it is really dealing with. The site carries most of the cost: fraud disputes, scraping, exhausted capacity and shared accounts.

Blocking every agent turns away real customers. Allowing every agent invites abuse. The sensible position is in between, and it starts with knowing what is arriving.

Our recommendation to anyone running a consumer website or public API in Malaysia has three parts. Start logging and classifying agent traffic now. Set a policy for each agent. Do not build your own delegation scheme while the standards settle.

You do not need Cloudflare to start

Cloudflare sells features for this, but the methods underneath are open. Four checks work on any web server, load balancer, API gateway or firewall. Use them in layers.

Check What it tells you How far to trust it
The name the request gives itself (the User-Agent, such as ChatGPT-User, ClaudeBot or PerplexityBot) Which agent it claims to be Weak. Anyone can copy the name
The operator’s published addresses Whether the claim comes from that operator’s network. OpenAI, for example, publishes its address list Medium. Lists change, so refresh them daily
A cryptographic signature (Web Bot Auth) Proof of which operator signed the request, checked against its public key Strong
Behaviour Automation that pretends to be an ordinary browser: speed, patterns, cloud-hosted origins Indicative, not proof

The first two are a matter of hours to days: a log filter and a scheduled job that pulls the published lists.

The third is the one to plan for. Web Bot Auth builds on a published internet standard for signing web requests, and you can verify signatures without being a Cloudflare customer:

  • Open-source libraries. Cloudflare publishes its verification code under an open licence, and it runs on any server.
  • Your existing firewall. AWS WAF and Akamai both describe support for verifying these signatures. One report says the AWS feature first applied only to CloudFront, so check coverage for your setup.
  • Your API gateway. Some gateways ship a ready-made policy.

What to log from day one

Ask your team to record, for every request:

  • Time, path and result
  • The name the client gives itself
  • The address and network it came from
  • Whether it carried a signature, and whether the signature was valid
  • The account or session it used
  • How fast that client is making requests

With those fields, a month of logs answers three questions: which agents visit us, what they do, and what it costs us.

The blind spot

Some agents drive an ordinary browser and announce nothing. To your logs they look like a person using Chrome.

Until signing becomes common, behaviour is the only view of them: many sign-ins to one account, a checkout completed faster than anyone can type, thousands of product pages read in order. These are the same signals that catch older kinds of abuse, and they are worth having either way.

Set a policy per agent

Treat agents as you would any other channel. Decide by agent and by action, according to how sure you are of who it is.

Tier Browsing and search Sign-in, booking, buying, paying
Verified and wanted. Valid signature or matching addresses, from an operator you choose to serve Allow, with generous limits Allow, with the same fraud checks as people, and a one-time code on high-value actions
Declared, unverified. Gives an agent name but nothing proves it Allow, with lower limits Challenge, or require a signed-in person to confirm
Undeclared automation. Behaviour signals only Limit or challenge Block, or force a one-time code
Known abuse. Forged identity, scraping, password guessing Block Block

Three rules keep this practical.

Decide by action, not by page. Reading a product page is low risk. Opening an account, redeeming a voucher or paying is not.

Write it down. One page, owned by a named person, so product, security and legal make the same call.

Review it monthly. New agents appear every quarter. Your logs will show which bring revenue and which bring cost.

Do not invent your own delegation scheme

Delegation is the hard part: how a customer’s authority passes through one agent, or several, without losing track of who is accountable. Cloudflare’s post says plainly that delegation across several intermediaries is not solved yet.

It is tempting to design your own token or header for “this agent acts for this customer”. We would not. The agents your customers use will not adopt it, and you will have to remove it when a standard arrives.

The standard is being written at the Internet Engineering Task Force. A Web Bot Auth working group is active, and its main specification became a working-group draft in September 2026. It is still a draft, and details may change. A separate proposal explores how a site can trust a request without learning who sent it.

What to do instead:

  1. Verify the signatures agents already send.
  2. Keep that logic in one place in your system, so the method can be swapped later.
  3. Ask one engineer to check on the working group each quarter.

Logs are personal data

Agent logs hold addresses and account identifiers tied to your customers. Under Malaysia’s amended data protection law, treat them accordingly. We covered what the amendments mean for software teams separately.

  • Involve your data protection officer early. Agree the purpose, which is preventing fraud and abuse, the fields kept, and who may read them.
  • Set a retention period. Keep detailed logs for a fixed time, and only totals after that. Pick a period your legal adviser can defend.
  • Say so in your privacy notice. State that you classify automated traffic for security.
  • Connect it to incident response. Password guessing or account takeover by automation can become a notifiable breach. Agent alerts should reach the same person as any other security alert.

Banks, insurers and e-wallet operators should fold this into their existing technology risk controls, not run it as a side project.

Who should act, and who can skip this

Act now if an anonymous or self-registered visitor can reach your site or API:

  • Online shops and marketplaces
  • Ticketing, travel and restaurant booking
  • Banking, insurance and e-wallets
  • Telco and utility self-service
  • News, content and classifieds
  • Any public API without strong client authentication

This matters much less for internal systems and business-to-business portals where every user signs in and every client is under contract. There, you already know who is accountable.

A 90-day plan

Days 1 to 30: see it. Add the log fields. Label requests by declared name and published addresses. Produce a first report: agents’ share of traffic, the top agents, what they request, and failed sign-ins and checkouts.

Days 31 to 60: verify it. Turn on signature verification through your firewall, gateway or the open-source library. Add behaviour rules for automation that does not declare itself.

Days 61 to 90: govern it. Approve the one-page policy. Apply limits and one-time codes on high-risk actions. Set the monthly review.

A question for the management meeting

Ask this at the next one: what share of our traffic last month came from AI agents, and which of them placed orders?

If nobody can answer, that is the first piece of work. It needs no new vendor and little money. It needs someone to own it.

Sources

This article is general information and is not legal advice.

Written by

Pikabyte Engineering

Pikabyte builds custom business systems, applied AI and secure cloud infrastructure from Malaysia. Talk to us.

Start a conversation

Want this applied to your business?

Tell us what is slowing the business down. You will talk to an engineer, not a salesperson, and get a straight answer on whether software is the right fix.