Consumer Protection Tuesday: How we leveraged AI to stop a popular Tesla Scam

By Coinbase • 4min read

Tl;dr: Scammers used an AI deepfake of Elon Musk to promote a fake Tesla crypto giveaway to a livestream audience of more than 500,000 people. In about two hours, hundreds of Coinbase customers tried to send millions to the scammers. Our AI-powered scam defenses stopped about 91% of that money, $4.7 million, from leaving customers' accounts, with no humans in the loop. Here's how it worked, and what to watch for.


Consumer Protection Tuesday: How to Spring Clean Your Digital Life

This post is part of a weekly Tuesday series at Coinbase about the latest consumer protection and security measures for crypto owners.

At Coinbase, we're on a mission to help update the financial system to make it safer and more secure. While under 1% of blockchain transactions are used for illicit activity, and cash remains the preferred medium for illegal transactions, crypto security is always a top priority. Coinbase maintains a robust compliance program, which includes Know Your Customer (KYC) checks, sanctions screenings, suspicious activity reporting, and strong law enforcement partnerships to detect and prevent illicit activity on our platform.

The Scam: A Livestream That Wasn't Real

During Tesla's "flying car" demo, scammers spun up a YouTube livestream of the event. It looked like the real broadcast. It featured a deepfake of Elon Musk, on-screen wallet addresses, and a QR code, and it promised to double any crypto sent to it.

More than half a million people watched.

The part that makes this hard: nobody was hacked. No Coinbase account was broken into, no password stolen, no credential phished. Hundreds of customers, many of them long-time users, decided on their own to send money to a stranger. The funds moved because the customer was convinced, not because an attacker forced their way in.

Why Scams Like This Are So Hard to Stop

Most fraud defenses are built to spot activity the customer didn't authorize. That's not what happened here. The customer was logged in, the customer pressed confirm, and the customer believed it was the opportunity of a lifetime.

Our systems can see risk signals in real time. They can't see intent. When someone is authenticated and approves a transfer, there is no way to tell a bad decision from a deception just by looking at the transaction.

So there was only one thing left to do: reach the customer in the middle of the decision and break the spell. Slow them down. Show them what they were actually looking at. Give them the specific reason this particular offer was a scam, not a generic warning they'd click past.

How Our Scam Defenses Work

We built a scam prevention agent that does exactly that. It has three parts:

1. Detection. Machine learning models score risky sends in real time, including sends tied to giveaway patterns, brand impersonation, and newly created wallet addresses. The goal is to catch the danger, not the customer.

2. Intervention. Once a send is flagged, a harness built specifically for scams takes over. It has tools to look up context, so it can respond to what the customer is actually telling it, and it holds a short, targeted conversation: it names the scam pattern, explains how the giveaway mechanics work, and asks the questions that make the story fall apart.

3. A confidence score on every reply. Each response the agent produces is scored before it reaches the customer, so the system knows when it's on solid ground and when a situation needs more friction.

What's notable is that the agent had never seen this scam. Deepfake giveaways tied to a live product launch weren't in its training. It recognized the shape of the scheme anyway, a promise to multiply funds, the urgency of an event that would end in hours, a transfer that could not be reversed, and it ran from the first flagged send to the last, with no humans in the loop.

Why Deepfakes Change the Math

Before, staging a scam like this took a crew, a studio, and weeks. Now it takes a livestream slot and a model. Deepfakes make convincing scams cheap, fast, and easy to relaunch under a new name the moment one is shut down.

That means our defenses have to improve at the same rate the attacks do. We're investing heavily in AI across Risk to make that happen:

  • Models that get better through AI-driven research, so detection keeps pace with new scam formats instead of waiting for the next wave.

  • AI agents that review risky sends across every type of scam, not just the ones that made headlines.

  • AI labeling of new scam cases, so our models learn from each attack sooner — often within hours instead of weeks.

What You Can Do

  • Nobody legitimate will double your crypto. An offer that requires you to send funds first to receive more back is a scam, no matter who appears to be making it.

  • Don't send crypto to an address or QR code shown in a video, livestream, or message. Anyone can put a wallet address on a screen.

  • If a send gets flagged, take the breath. Slow down, read why, and talk to someone you trust before you confirm.

  • Verify somewhere else. Close the video, and check the company's official account or website directly instead of using a link or address from the post.

  • If you think you've been targeted, report it. Contact Coinbase Support through the app or Coinbase Help right away.

Coinbase Will Never

  • Ask you to send crypto to claim a prize, enter a giveaway, or verify your wallet.

  • Promise to double, triple, or multiply crypto you send.

  • Ask you to rely on a QR code shown in a livestream as the only way to take part in a promotion.

  • Pressure you to act before a livestream ends or a countdown runs out.

Recent stories