← All articles
Agent security • Field note

An account takeover is a path, and it runs through support

A security vendor now maps attacks by the person they start from. In a customer business, that person is often talking to your support desk, one routine request at a time.

David Soden • 5 min read • 13 November 2026
A support agent in a headset taking a call at his desk
Support exists to help people who can't get into their account. So does an attacker's script.

What happened

ReliaQuest added GreyMatter Attack to its GreyMatter security platform. A security analyst types a question in plain language and gets back a map of how an attacker could move through the company, starting from a user identity, an access point or a known technique. The team then checks whether each path really works, and the findings go to ReliaQuest's agents, which write detections, start responses and apply controls.

This is a product launch from a security vendor, aimed at a security operations team, and it will be on show at Black Hat USA 2026. One detail is worth lifting out of that setting. The paths start from a person and what that person can reach, rather than from a server.

Why this matters to your business

Follow a path that starts with a person through a customer business and it often runs through the support desk. Support is there to help people who are locked out, and a locked-out customer and an attacker ask for the same things. Change the email on file. Reset the password. Update the phone number. Send the replacement card to a new address. Each request is routine, and each one, granted to the wrong person, hands over the account.

The path rarely shows up in one contact. Someone asks chat to change the email on Monday. A different voice phones on Wednesday to reset the password, and the reset link goes to the new email. On Friday the payout account or delivery address changes. Every agent who handled a step followed the script, because each step on its own was allowed. The pattern only appears when you read the three contacts together, and nobody on the floor sees all three.

The bill lands in familiar places. You owe the real customer their money back, and you keep a customer who no longer trusts you with their card. If you have given an AI agent on chat the power to change account details, it now handles those requests faster and around the clock, which helps the attacker as much as anyone.

A line of footprints crossing fresh snow on a dark road at night
One footprint tells you little. The line of them tells you where someone is going.

Why this is a CX-Builder use case

The useful idea in the launch is to judge a request by the path it completes. For a support team, that means writing down the handful of sequences that end in a takeover in your business, then checking every sensitive request against the account's recent history before it goes through.

CX-Builder can build that check as one flow that sits in front of every account change, whether a person or an AI agent is making it. The flow reads the account's recent events across chat, phone, email and the account itself, and compares the request with the paths you wrote down. A plain request goes straight through. One that completes a known path, such as a password reset three days after an email change from a new device, stops at a human-in-the-loop step. It goes to someone who can call the customer back on the number you had last month.

The paths live in a document the agent retrieves from, so your fraud lead can add a new one the week it shows up without anyone touching the flow.

A hand holding a small brass padlock against a dark background
The check belongs in front of the change, not in the review after it.

What this looks like if you build it

An agentflow starts whenever someone requests a sensitive change. Tool calls pull recent account events from your CRM and help desk through their APIs. An agent compares the request with the path library, held in a vector store, and returns a structured verdict: allow, hold, or decline, with the path it matched. Holds go to an approval step, and every verdict is logged with the events behind it. Self-hosting keeps account history and your list of known paths on your own servers, which matters when that list is a guide to your weak spots.

Start with the paths you already know. Your fraud team and senior agents can name the last few takeovers and how each one played out. That list is the first version of the library.

The takeaway

Pull the last five account takeovers your team dealt with and write down each step, contact by contact. If any of them crossed more than one channel or more than one agent, your support desk can't see that path today. Put a check against those paths in front of email, password and payout changes before you give an AI agent the power to make them.

All articles Install CX-Builder View on GitHub