How Coinbase AI Stopped $4.7M In A Musk Scam It Had Never Seen

Broadcast camera aimed at a flying-car prototype on an empty launch stage

A fake Tesla livestream showed an Elon Musk deepfake promising to double viewers’ crypto. On November 17, 2026, Coinbase said its AI system had blocked about $4.7 million that customers tried to send to the scammers, roughly 91% of the attempted funds. Many of those customers had used Coinbase for years. They logged in and approved the transfers themselves. Coinbase had to identify a scam in transactions its customers wanted to make, although its agent had never been trained on deepfake giveaways tied to a live product launch.

More than 500,000 people watched the fraudulent YouTube stream during Tesla’s “flying car” demonstration, Coinbase said. The video imitated the real broadcast and displayed wallet addresses and a QR code. Over about two hours, hundreds of Coinbase customers tried to send crypto to those addresses.

Coinbase described its response in three steps:

  1. Risk models scored outgoing sends in real time for patterns including giveaway promises, brand impersonation and newly created wallet addresses.
  2. AI agent took over flagged sends, explained the suspected scam and asked customers questions about the offer.
  3. Confidence check scored the agent’s reply before the customer saw it and determined whether the transfer needed more friction.

The agent identified the promise to multiply funds, the pressure of a live event and a transfer that could not be reversed, Coinbase said. The system could then intervene without a person reviewing each send.

Our systems can see risk signals in real time. They can’t see intent.

Coinbase

Roughly $460,000 in attempted transfers were not blocked. Coinbase did not say how much reached the scammers or when the fake stream ran. Most of the attempted funds stayed in customers’ accounts despite their approval of the sends.

Coinbase said it is expanding AI review of risky sends beyond high-profile scams. It is also using AI to label new scam cases so its risk models can learn from attacks within hours instead of weeks.