Coinbase said on Tuesday 17 November 2026 that its automated scam defences stopped $4.7 million in customer transfers headed for a fake Tesla livestream, where a deepfake of Elon Musk promised to double any crypto sent. That sum, the exchange says, was about 91% of what its customers tried to send. The account appeared on the company’s blog and was relayed by The Block, Bitcoin Magazine and Bitcoin.com News, which quotes the exchange’s post on X: “Our AI-powered scam defenses stopped $4.7 million from leaving customers’ accounts.” Every figure is Coinbase’s own, and no outside party has checked any of them.
The stream ran on YouTube last month, during Tesla’s “flying car” demonstration, according to The Block, which gives no exact date. It imitated Tesla’s official broadcast. Coinbase says more than 500,000 people watched it and that, within about two hours, hundreds of its customers tried to send millions of dollars to the addresses shown on screen. Many of them, it adds, were “long-time users”.
The post leaves out as much as it gives. Coinbase does not say how much money got through, how many customers were involved, or whether the 91% counts transfers or dollars.
Faked faces turn up in much larger thefts too. Five days ago we reported on the FaceStealer case, in which the FBI says North Korea stole $400 million with fake Zoom calls built on real webcam clips.
How the deepfake Musk giveaway scam worked
The method is old and the production is new. A channel broadcasts what looks like a company event, with a familiar face on stage. Wallet addresses and a QR code sit on the screen. The pitch, in this case, was that any crypto sent during the event would come back doubled. The offer ends when the stream does, so viewers have a couple of hours to decide, and a transfer made on a blockchain cannot be called back once it confirms.
What made this one work was the frame around the fake Musk: a copy of a real Tesla broadcast, timed to a real product demonstration that people were already looking for. Our guide to fake airdrops and giveaway scams describes the same “send and get double” pattern in its older forms.
No account had to be broken into. “The part that makes this hard: nobody was hacked,” Coinbase wrote, as quoted by The Block. “No Coinbase account was broken into, no password stolen, no credential phished.” Customers logged in with their own credentials and approved the sends themselves, which is why the usual defences against account theft, from passwords to two-factor codes, had nothing to catch.
Bitcoin Magazine quotes the company on the limit this puts on any filter: “Our systems can see risk signals in real time. They can’t see intent.”

A risk score first, then an AI agent that questions the customer
As The Block and Bitcoin Magazine describe it, the system works in two stages. Machine-learning models score outgoing transfers in real time, looking for giveaway patterns, brand impersonation and destination wallets created very recently. A transfer that scores as risky is passed to an AI agent built for scams. The agent pulls the context of the send, names the kind of scam it resembles, explains how such a scheme is built and asks the customer questions meant to test the story they have been told.
Each reply from the agent is scored for confidence before the customer sees it, Bitcoin Magazine reports, and that score decides when to put more obstacles in front of the transfer. Coinbase says the whole episode, from the first flagged transfer to the last, was handled “with no humans in the loop”.
The company also says the agent had not met this variant before. “Deepfake giveaways tied to a live product launch weren’t in its training,” it wrote, according to The Block. By Coinbase’s account it recognised the parts anyway, among them a promise to multiply funds and the urgency of an event that “would end in hours”.
Coinbase has not said what the customer actually sees: whether a flagged transfer is held, delayed or only interrupted by a conversation, whether a customer who insists can still send, or which coins were involved. The announcement landed the same day as another AI launch from the exchange, the Coinbase trading app inside ChatGPT.
What Coinbase does not say about the 91%
The headline figure has no stated basis. The Block writes of “about 91% of the attempted transfers”; Bitcoin Magazine reads it as roughly 91% of the money customers tried to send. Those are different measures. If a few large transfers went through while many small ones were stopped, a share counted in transactions would flatter the result, and the reverse is just as possible.
Bitcoin Magazine did its own arithmetic. If 91% is a share of value, customers tried to send about $5.2 million and roughly $460,000 still left their accounts. That is the outlet’s estimate under that assumption, not a number Coinbase published. The company itself speaks only of “millions” attempted.
Coinbase has not said either:
- How much reached the scammers, and whether any of it was recovered or refunded.
- How many customers tried to send, beyond “hundreds”, and how many went ahead after a warning.
- How many legitimate transfers were flagged by mistake during the same period.
- Which coins and networks were used, and which addresses received the money.
- Whether the stream was reported to YouTube or to the authorities, and whether YouTube removed it.
The three outlets name no Coinbase executive and carry no comment from Tesla, Musk or YouTube. The viewer count is also the company’s: it is not said whether 500,000 means people watching at once or total views, and scam streams have a record of inflated audiences.

What an exchange can block, and what it cannot
Coinbase can only examine transfers that leave accounts it holds. A viewer who scanned the same QR code with a self-custody wallet, or who sent from another platform, never met this filter. A wallet under your own keys signs what you tell it to sign, and no agent steps in to ask why.
So the $4.7 million describes one company’s customers over two hours. It says nothing about what the stream collected in total. Because no destination addresses have been published, nobody outside Coinbase can look at the blockchain and count what arrived.
The nearest precedent dates from 15 July 2020, when hijacked Twitter accounts, Musk’s among them, posted a bitcoin doubling offer. Coinbase was “blocking any attempted payments to the scammers’ address”, Fortune reported that day. By 6 p.m. New York time the address had still received just under 12 BTC, about $109,000 then, part of which may have come from the scammers themselves. That was one known address put on a list. The system Coinbase describes in 2026 has to react to behaviour, before anyone knows the address.
Deepfake scam losses in figures
None of the public statistics isolates giveaway streams, but the reports published this year all point the same way on impersonation and AI. Each figure below comes from the named source and covers the stated period only.
| Source | Period | Figure |
|---|---|---|
| FBI Internet Crime Complaint Center (IC3), annual report released in April 2026 | 2025 | 181,565 crypto-related complaints and $11.366 billion in reported losses, up 22% on 2024; 22,364 complaints with an AI element, $893 million in losses |
| US Federal Trade Commission, 27 April 2026 | 2025 | $2.1 billion reported lost to scams that began on social media, eight times the 2020 figure, of which $1.1 billion in investment scams (scams of every kind, crypto or otherwise) |
| Chainalysis, 13 January 2026 | 2025 | At least $14 billion received by crypto scams on-chain; impersonation scams up 1,400% on 2024, and AI-linked scams 4.5 times more profitable than the others |
| TRM Labs, 2026 AI-in-Crime Adoption Index, 17 August 2026 | 2026 to mid-August, against all of 2025 | Reported deepfake-scam losses already 263% above the full-year 2025 total (TRM’s own data) |
| DFRLab, 25 September 2024 | 18 June 2024 | One deepfake Musk livestream on a hijacked YouTube channel: up to 81,000 concurrent viewers, about $50,537 collected |
The FBI figures are those The Block took from the IC3 report on 12 April 2026; the FTC, Chainalysis and TRM Labs published theirs directly. TRM adds a detail on cost. One vendor it tracked sold a face model for roughly $500 and a one-year deepfake package for about $3,000.
The 2024 case studied by the Atlantic Council’s DFRLab gives a sense of scale. That stream looped a 200-second clip from a Tesla event, pointed its QR code to a site named tesla-bitcoin.com and took in about $50,537 in bitcoin, ether and dogecoin. The $4.7 million Coinbase says it stopped at a single stream is about 93 times that sum. The two numbers come from different sources and methods, so the gap is an indication, not a measurement.
Signs of a fake giveaway stream
The pattern Coinbase says its agent picked out is one a viewer can check without any software. What follows is general guidance, not Coinbase’s wording.
- A promise to send back more than you send is the scam itself. No company or founder multiplies crypto sent to an address.
- A wallet address or QR code on screen during a live event, with a countdown or an offer that ends within hours, is there to stop you from checking.
- Look at the channel, not the badge. In the 2024 case DFRLab found a hijacked channel with 420,000 subscribers and a YouTube verification mark, renamed “Tesla” for the occasion, with comments disabled.
- If your exchange interrupts a transfer with a warning, read it before clicking through. Once the transfer confirms, nobody can reverse it.
- Report the stream to the platform, then tell your exchange and the police or fraud agency of your country if you have already sent money.
Our guide on how to avoid crypto scams lists the other common types and where to report them.
Next, Bitcoin Magazine reports, Coinbase plans to extend AI review of risky sends to all scam types and to use AI to label new cases, so that its models learn from an attack within hours instead of weeks. It has given no date for either step, and no figure yet on how often the agent stops a transfer that was legitimate.
