The Web Form Attack That Can Run Up Your Voice Agent Bill Overnight

My AI got attacked. By the time I caught it, the bill was over $137, and the only reason it was that low is I caught it literally as it started. Others in my community weren't so lucky. One person lost $18,000.

If you have a voice agent that calls people back after they fill out a form on your website, or you're thinking about getting one, you need to know about this. I have a bachelor's degree in cybersecurity, so I look at this as both a builder and someone trained to spot vulnerabilities.

How the attack works

From what I can tell, these attacks target outbound agents, specifically the kind that call someone as soon as they fill out a form on your site. That's exactly the setup I had.

Bots fill out the form over and over with random names and fake company emails. The one thing in common was the phone numbers: rotating UK numbers. And those numbers aren't connected to a person. They're connected to a recording that sounds just human enough to keep your agent talking. "Hi, I have a question." "I can't hear you, I have a sound problem, wait a minute." One kept asking my agent if it spoke Arabic. Meanwhile the meter keeps running.

My best guess on their setup is a basic script spamming forms, cheap internet phone numbers bought in bulk for around a dollar each, and recordings played on repeat. It probably costs them 25 to 100 bucks max. It can cost you thousands overnight.

What it actually cost people

I was literally in a meeting talking about this exact attack. An hour later I pulled up my demo agent to show a friend, something I normally only check about once a week, and my concurrency was maxed out. 20 out of 20 calls happening at the same time. That's not normal. I detached the agent from the number right there.

It was climbing about $10 a minute. If I hadn't caught it, that's $600 by the end of the hour and around $5,000 by the next morning.

One builder in my community lost $450. Another lost $600. Someone he knows got charged $18,000. A friend of mine lost over $300, and the scary part is he had CAPTCHAs and spam filters in place. The bots got around them anyway.

Why do they do it? Honestly, I don't know. I don't see any profit in it for them. My best guess is trolling or harassment. Either way, it's crazy expensive for whoever owns the agent.

Old defenses aren't enough

My friend's story is the lesson. CAPTCHAs and honeypots are the security of yesterday. A honeypot might slow an attacker down like wading through mud, but it's not going to stop them from getting over the wall anymore. We can't expect yesterday's cybersecurity to stop today's AI attacks.

The three protections I use now

First, only call US numbers. The numbers on the platform I build on can now call internationally, and every one of these attacks came from international numbers. That doesn't mean it can't happen with US numbers, but it closes the door they've been walking through.

Second, and this is the big one, a confirmation text. When someone fills out the form and asks for a call, the agent texts them first: hey, this is Mark's agent, just confirming, reply yes if you still want this call. A bot spamming a form can't confirm that text. I don't see a way around it, though I could be wrong.

Third, a cooldown. Make the system wait 60 seconds between calls, or cap it at a couple of calls per minute. Even if something slips through, it can't max out 20 lines at once.

What this means if you own the agent

Don't wait for the bill. If a voice agent is wired to a form on your website right now without protections like these, seriously consider taking it down until they're in. Don't wait until you wake up to a bill of $5,000 or more.

And whoever builds your agent should be able to tell you exactly what stops a bot from making it dial a pile of fake numbers overnight. That's a fair question to ask before anything goes live.