Fake YouTube tutorials promising AI-built crypto arbitrage bots led 224 victims to deploy and fund their own malicious smart contracts, draining 274.60 ETH between Feb. 12 and Aug. 11, according to a Sept. 14 analysis by blockchain intelligence firm TRM Labs.

TRM identified 234 victim-deployed contracts forwarding funds to six operator-controlled collection addresses. The firm valued the stolen ETH at approximately $517,205 at the time of the transfers and put the median loss at 1 ETH per victim.

The scheme breaks the template of conventional drainer campaigns. It did not depend on a victim signing a malicious token approval or visiting a spoofed domain. Each victim picked a tutorial, copied code, deployed a contract and funded it from their own wallet — a sequence that makes every transaction look self-directed to wallet security systems.

How did the fake compilers work?

TRM found nine nearly identical YouTube tutorials presented as the work of separate creators. The videos promised viewers an arbitrage bot built with Claude and directed them to compiler sites controlled by the operators, some styled to resemble the widely used Remix development environment.

In one variant analyzed by TRM, a background script discarded the source code the victim pasted and fetched a different contract from the operator's server. The clean code displayed on screen never reached the blockchain.

The substituted contract accepted deposits and forwarded any balance above 0.05 ETH to the operator when the victim pressed Start or Withdraw — the exact actions the tutorial instructed them to take. TRM said the contracts contained no arbitrage logic and no AI functionality; the Claude branding served purely as marketing.

What happened after the first drain?

One compiler site attempted to extract a second payment after the initial theft. It displayed an invented "gas nonce liquidity" error and told the victim to add another 50% of the original deposit, up to 1 ETH. TRM noted the term is not an Ethereum concept and the message was engineered solely to prompt another transfer.

The nine videos remained online as of September, with a cumulative 310,474 views, according to TRM's count. The firm also documented earlier versions of the scheme that used ChatGPT as the lure in 2025 — evidence that operators can swap the AI branding without altering the underlying theft mechanism.

The operational takeaway for security teams is uncomfortable: on-chain behavior alone cannot flag these attacks. Deployment and funding transactions originate from victim wallets at the victim's initiative, bypassing the approval-anomaly heuristics that catch conventional drainers. Detection now has to move upstream — to the tutorial content, the compiler infrastructure and the six collection addresses TRM has already mapped. As long as those videos stay live and accumulate views, the operator addresses and their off-chain funnels remain the practical intervention points for exchanges and analytics vendors tracking the flow of stolen ETH.