AI agent fraud
Account takeover, fake accounts and checkout abuse can all arrive through an agent. The defense is the same as it's always been: know who you're really dealing with. The agent just makes it easier to forget to ask.
AI agents are already on your site. Some read your pages so ChatGPT or Claude can answer a question. Some compare prices, check availability or fill in forms. Some add items to a cart. An agent is a middleman: its name proves nothing, anyone can pretend to be it, and the same agent can carry a loyal customer one minute and a fraudster the next. Decide on what it's trying to do, and who it's doing it for.

Definition
AI agent trust management is how a business identifies the AI agents and bots reaching its website, decides what each one may do, and checks who is behind the request. It covers declared crawlers, signed agents, and automation that hides or impersonates, and it applies different rules to browsing, login, signup and checkout.
But it's one agent among many, and new ones arrive every month. This guide covers all of them, with Muse as a worked example.
Five ways AI agents show up on a site. Pick yours and jump to the part of this guide that answers it.
Reading your content and prices.
Comparing, buying, booking.
Logging in, applying, changing details.
Scrapers and attackers posing as agents.
Customers or automation?
01 · What's on your site
Not all AI traffic looks the same, and it doesn't all deserve the same answer. There are four kinds of AI bot traffic.
GET /products/hiking-boots
User-Agent: Meta-ExternalAgent/1.1The bot says what it is in its user agent, like a search or AI crawler.
Easy to see. Easy to copy.
GET /products/hiking-boots
Signature-Agent: [agent key directory]
Signature: sig1=:[verified]:The agent signs its requests, so you can check cryptographically that it comes from who it claims.
Proves the agent. Nothing more.
GET /products/hiking-boots
User-Agent: Mozilla/5.0 … Chrome/…
(no signature)A normal browser, a normal user agent, no signature. Many of today's agents work this way.
Invisible to a name check.
POST /login
User-Agent: ExampleAgent/1.0
(claim, not proof)Traffic that claims to be an agent it isn't.
The moment you allow an agent, it becomes worth dressing up as.
Kasada network
Across the sites Kasada protects, AI agent traffic is climbing fast, but it's still a small slice of the total. That's the point: set the rules while it's small.
Verification
robots.txt is a request, not a lock, and a user agent can be copied in a line of code. Signed agents are real progress, but a valid signature only tells you which agent it is.
# robots.txt
User-agent: Meta-ExternalAgent
Disallow: /search/
# Read by crawlers that choose to listen.
# Ignored by anything that doesn't.Well-behaved crawlers like Meta-ExternalAgent read it. Agents that act for a person, and anything that doesn't want to be stopped, often don't.
A user agent is a suggestion, not an identity.
Signature-Agent: [agent key directory]
Signature: sig1=:[verified]:
# Which agent: answered
# Who it's acting for: not answeredA verified agent can still carry a stolen password, and one agent's reputation is shared by millions of people.
Identity without authorization is just better labeling.
The signature answers it.
Your policy answers it. See the policy check below.
The account behind it answers it.
02 · AI agent access control
Allow or block is the wrong unit. These are starting points, not recommendations: your rules depend on your customers, your risk appetite and what each agent brings you.
Six questions, one journey at a time
You leave with a list of the agents you see and how each is verified, a rule for each step of the journey, an owner and a review date.
Run the check with our teamWhich agents reach this journey, how much, and on which endpoints? Your analytics may undercount them, so check your bot-management or edge logs.
A user agent anyone can copy, an IP range that's hard to keep current, or a signed request? Rate each agent you see.
How would you spot traffic claiming to be an agent it isn't, or an agent hiding as a person? A CAPTCHA won't settle it: an agent may solve it, fail it, or hand it back to the person.
For each step (browse, search, add to cart, log in, create an account, check out), decide what you'll allow.
Ask Marketing which agents send real customers before you block them.
Allow one agent on one journey for a fixed window, keep checkout out of the first test, and agree the turn-off criteria in advance. Name an owner and a review date.
03 · Who's behind the agent
When an agent logs in, applies or buys, it usually acts inside a real person's account. That account already has a history. That's a record, not a guess. So that's where the decision belongs: verify the agent, then check the person behind it, and act on what you find.
POST /login, via an AI agent.
Which agent is this, and is it really that agent?
A stolen password looks different from its owner, even when both arrive through the same agent.
Agents don't create new kinds of fraud so much as new routes for old ones.
Account takeover, fake accounts and checkout abuse can all arrive through an agent. The defense is the same as it's always been: know who you're really dealing with. The agent just makes it easier to forget to ask.
For customers using Account Intelligence, Kasada scores the person behind a request when they share identity, such as an email at login, signup and sometimes checkout, including when they arrive through an agent. We're working to extend this.
04 · Example
Muse is the agent everyone is asking about, so here's the framework applied to it.

Muse is Meta's AI agent. It browses, compares and shops for people from Meta's apps.
Mostly nothing you can check easily. Most Muse traffic is unsigned, Meta doesn't publish an IP range, and requests arrive with a standard Chrome user agent from cloud machines. Meta has started signing a share of Muse requests with Web Bot Auth, and that may grow.
Block it outright and you turn away real customers who chose to shop that way. Allow it on name alone and you let in everyone who uses it, and everyone who can pretend to be it.
Applying the check
Browsing and product pages: usually fine to allow. Login, signup and checkout: a higher bar until you understand how your customers use it. Every step: check the account, not just the agent.
05 · For Marketing teams
Marketing teams are being told to get found by AI agents. Fair. But before you celebrate the AI traffic in your reports, ask what's in it: customers using agents to research and buy, or automation that looks like demand. Get it wrong and you fund the wrong campaigns.
Does GA4 show AI agent traffic? Partly. It can show visits referred from AI assistants. But an agent browsing for a person with a normal browser looks like an ordinary visit, and traffic that never runs your analytics script doesn't appear at all.
The best agent policies are written by Security and Marketing together. Go deeper with Forrester's best-practices report.

It reads signals from the browser itself, not a user agent anyone can copy. No CAPTCHAs, so real people and the agents you allow never solve a puzzle.
Allow, read-only or block each agent, or set rules by endpoint.
We add new agents to the Verified Bots and AI Agent directory as they appear, starting with the most widely used.
With Account Intelligence, when the person shares identity such as an email. We're extending this.
Kasada is a Leader in The Forrester Wave™: Bot and Agent Trust Management, Q2 2026.

On-demand webinar
Our on-demand session on deciding what AI agents may do.
FAQ
Not across the board. Blocking turns away customers who use agents; allowing on name alone lets in anyone who can pretend to be one. Decide per action: allow browsing, hold login, signup and checkout to a higher bar, and check the person behind the agent.
Start with robots.txt for crawlers that respect it, such as declared AI and search crawlers. For agents and automation that ignore it or hide their identity, you need bot detection that recognizes them from their behavior and browser signals, not their name.
Only the ones that choose to follow it. robots.txt is a request. Agents acting for a person, and anything trying to avoid detection, can ignore it.
It depends on what you publish and what you get back. Some businesses block training crawlers but allow AI search and answer crawlers that can send them visitors. Make it a deliberate decision per bot, and review it as new ones appear.
Not by its name. A user agent can be copied in a line of code. Detection has to look at the browser and the behavior behind each request, and separate signed agents, undeclared agents and impersonators. Kasada grades each agent by what it can prove and flags requests that claim an agent's name but fail verification. See AI Agent Trust.
Yes. Anything you allow becomes worth impersonating. That's why a name or user agent should never be enough to let something in.
Mostly no. Meta has started signing a share of Muse requests with Web Bot Auth, but most traffic is still unsigned, there's no published IP range, and it uses a standard Chrome user agent.
Not reliably. An agent might solve a CAPTCHA, fail it, or hand it back to the person halfway through checkout. Kasada doesn't use CAPTCHAs; it detects automation from signals in the browser.
Let agents do what helps customers, like browsing and comparing, and hold the steps that move money or accounts (login, signup, checkout) to a higher bar. Verify the agent, then check the person behind it.
Run the policy check with our team. Bring one journey; we'll work through the six questions with you.
Watch “AI agent trust: what to allow, what to block, and what to prepare for.”
Forrester's best-practices report on telling trustworthy traffic from AI agents, bots and scrapers apart.