Nine YouTube tutorials promising to teach viewers how to build a crypto arbitrage bot with Claude ended up draining 274.6 ETH, roughly $517,000, from 224 people, according to a Sep. 14 report from blockchain intelligence firm TRM Labs. The scheme is notable less for the amount than for the mechanics. No phishing links, no wallet approval requests. The victims built the trap themselves, one step at a time, and signed every transaction with their own hands.
Each tutorial presented the same pitch: copy this code, open this compiler website, connect your wallet, deploy the contract, and fund it. The videos used AI-generated hosts and voiceovers, which gave them a uniform, professional look and let the operators publish them under nine different creator identities. The promised bot would exploit price differences between trading venues around the clock, generating passive income from the volatility that crypto markets are known for. The contract the backend actually served did something else entirely.
The compiler websites were the pivot of the scam. Some copied the design of Remix, the browser-based development environment that most Ethereum developers genuinely use for learning and prototyping. When a victim compiled their “trading bot,” the site’s backend swapped in a malicious contract variant designed to drain whatever the user deposited. Because the user compiled it, deployed it, and funded it through their own signed transactions, each step looked like a legitimate development workflow rather than an attack. On-chain, nothing about the sequence raised flags.
Why the scam worked
Wallet security warnings have trained users to distrust one thing above all: a website asking for a token allowance or a signature they do not understand. This campaign never asked for either. TRM found no arbitrage system and no AI function in the malicious contract variant it examined. The only real technology was the video production, and that technology was convincing enough to carry the whole operation.
The firm traced 234 contracts deployed by victims, more than one per person in some cases, with the stolen funds flowing to six collection addresses. The median loss was 1 ETH, which says the scheme took from many small holders rather than landing a single whale. That distribution is consistent with the tutorial format, which skews toward beginners with modest balances and no development experience to recognize that the code they are compiling makes no sense. A developer reading the contract would have caught the drain function in minutes. The target audience was precisely the audience that cannot read it.
The AI wrapper is the story
Attaching Claude’s name did two things. It borrowed credibility from a widely known AI product, and it matched the current wave of interest in AI trading tools, where the audience already expects that a bot might genuinely work. TRM’s report does not suggest Anthropic had any involvement. The name appears as bait, the way fake tokens borrow brand names during bull markets.
It still signals where fraud is heading. Generative tools have cut the production cost of convincing content to near zero, and crypto remains the payment rail where stolen funds are hardest to claw back once they leave the victim’s wallet. The combination produces campaigns that look like independent creators sharing knowledge and are, in fact, one operation wearing nine faces. The nine identities also provide redundancy: when a video gets reported and removed, the others keep working. Expect more of this shape, not a malicious link but a malicious process the victim executes with full consciousness at every step.
What to tell people who might see one
The tells are visible once you know them. A trading bot tutorial that requires deploying your own contract from a compiler site chosen by the video author is not a normal tutorial. Real development environments do not need you to fund a contract to make money for you, and no genuine arbitrage strategy gets shared in full on YouTube, because arbitrage edge dies the moment it is copied at scale. Code served by a website can be swapped at any time; the code shown in the video proves nothing about what the compiler returns to your browser.
The practical rule is blunt: never deploy a contract from a site a video sent you to, and never fund one. Anyone who did, thinking they had built a bot, can check the contract they deployed on a block explorer. If it sends ETH to addresses they do not control, it was never a bot. Recovery is unlikely in practice, but reporting the addresses helps firms like TRM cluster the campaigns, link them to earlier operations, and track the operators as they rebuild under new identities.
Security teams and educators also have an adjustment to make. Warnings built around phishing links and approval popups do not describe this attack. The new warning is about provenance: who controls the tooling between you and the chain. That covers compiler sites, code snippets pasted from chat apps, and any “verify it yourself” step where the verification tool itself is controlled by the attacker.
The last line of defense is skepticism about returns. An arbitrage bot that works does not need your 1 ETH to function, and whoever has one does not teach strangers how to replicate it on YouTube for free. The video promised passive income and delivered an empty wallet. Until that message reaches the tutorial-watching audience, campaigns like this one will keep finding takers, and the $517,000 figure will read as a modest quarter for whoever ran it.