Markets
BTC $56,260 -8.49% ETH $1,560 -9.57% BNB $514 -9.34% XRP $0.8765 -8.81% SOL $66.04 -8.41% LINK $8.15 -7.10% BTC $56,260 -8.49% ETH $1,560 -9.57% BNB $514 -9.34% XRP $0.8765 -8.81% SOL $66.04 -8.41% LINK $8.15 -7.10% TOTAL $1.98T BTC.D 57.3%

Blockchain Technology

Coinbase Says Its AI Agent Blocked $4.7 Million of a Deepfake Musk Giveaway

Customers were logged in and approved the transfers themselves, which is precisely the pattern that account-security tooling is not built to catch.

Coinbase says an AI-powered scam detection system stopped $4.7 million, about 91% of the funds targeted by a fraudulent Elon Musk crypto giveaway, from reaching the scammers’ wallets, with no human intervention at any point in the response.

The scheme ran through a fake YouTube livestream that appeared to show a Tesla flying car demonstration. It used an AI-generated version of Musk, displayed wallet addresses and a QR code, and promised viewers their money would be doubled if they sent cryptocurrency. More than 500,000 people watched.

What made the incident unusual is that nobody’s account was broken into. Customers were logged in, had access to their funds, and approved the transfers because they believed the giveaway was real. Conventional security controls look for unauthorised access or unusual account behaviour, and neither applied here.

Coinbase’s answer is an agent that assesses transfers in real time using machine learning models trained to spot giveaway framing, brand impersonation and payments to newly created wallet addresses. When a transfer is flagged, the agent opens a targeted conversation explaining the specific scam pattern rather than issuing a generic warning. Each response it generates carries a confidence score before being shown to a customer, which the company uses to judge when it has enough certainty to intervene.

The agent had never encountered this particular scheme; deepfake giveaways tied to live product launches were not part of its training. It still recognised the familiar markers, including promises to multiply holdings, time pressure and the irreversibility of the transfer, and handled the flagged attempts from the first to the last without a person stepping in.

“Before, staging a scam like this took a crew, a studio, and weeks. Now it takes a livestream slot and a model,” Coinbase said, adding that defences have to improve at the same rate as the attacks.

The broader trend is documented. A Gartner survey of 297 senior cybersecurity leaders, conducted between March and May 2026, found 41% of security chiefs had dealt with at least one social engineering incident involving an employee audio deepfake in the previous year, while 36% reported a video deepfake incident. In June, Japanese crypto advocate Mai Fujimoto was tricked on a Zoom call by a deepfake impersonating an acquaintance.

Share this article

Viber uses viber://forward?text= when the app is installed.