An AI trading bot scam has stolen 274.6 ETH, about $517,000, from 224 victims who thought they were building their own automated arbitrage system, according to blockchain intelligence firm TRM Labs. The scheme inverted the usual phishing playbook: instead of asking users for wallet permissions, nine near-identical YouTube tutorials walked victims through deploying their own malicious Ethereum contracts, using branding around Anthropic's Claude AI to sell the illusion. Every onchain action looked legitimate because the victim initiated it. That is what makes this campaign a template worth studying.
How the AI trading bot scam turned victims into deployers
The videos presented themselves as educational guides, fronted by AI-generated virtual hosts and voiceovers, each promising to help viewers create a fully automated crypto arbitrage bot. Viewers were told to copy code and open a compiler website chosen by the presenter, some of which mimicked the design of Remix, the widely used browser-based development environment for Ethereum contracts. Victims connected their wallets, compiled what appeared to be trading software, and deployed the result. Because users initiated and approved each step, the transaction pattern looked nothing like a conventional attack that requests a token allowance or a suspicious signature.
The trick lived in the backend. In the variant TRM examined, the compiler site ignored the source code victims pasted into the browser window and instead retrieved a separate contract from a server the operators controlled, preparing the replacement for deployment. The clean code on screen was never the code that went onchain. Users believed they were funding a bot that would exploit price differences between venues, but TRM found no arbitrage system and no AI function in the deployed contract at all. The moment its balance exceeded 0.05 ETH, the contract transferred everything to an operator address the next time the user pressed either the Start or the Withdraw button. Both buttons did the same thing: steal the deposit. Stolen funds flowed to six collection addresses.
Why wallet defenses never fired
Traditional crypto phishing leans on copied domains, poisoned search results, or prompts for broad token permissions, and defenses are built for that shape. Blocklists flag known malicious addresses, and wallet simulations try to warn users about dangerous transactions. This campaign defeated both by moving the deception one step earlier. Every victim became the deployer of a freshly created contract, so no blacklist contained its address, and no seed phrase or allowance was surrendered. A wallet can accurately show its owner deploying a contract, funding it, and calling one of its functions, while lacking the context to know that the tutorial and the compiler had lied about what the code does.
The damage profile is unusually democratic. TRM calculated a median loss of 1 ETH per victim, meaning the total did not hinge on a whale. Some of the 224 victims created more than one contract, which is why 234 deployments were traced. The numbers suggest the audience was ordinary retail users chasing a plausible AI-era windfall, the same population that gets hit by fake trading app scams in other formats, and the same supply of targets ransomware operators hunt when they want payment in crypto, as in the Revolut extortion case.
The AI wrapper is the story
Strip away the blockchain and this is an old con: a tutorial teaches the mark to build the trap himself. What is new is the packaging. AI branding provides credibility, AI-generated hosts manufacture independent-creator authority at zero cost, and AI anxiety creates the fear of missing the arbitrage wave. The same pattern is visible across the industry's threat reports: drainer infrastructure now operates as a service, splitting stolen funds among connected addresses automatically, and recruitment-laced malware targets developers directly. The industry response has been institutional, with security alliances sharing drainer intelligence across MetaMask, Phantom, WalletConnect and Backpack, and reported drainer losses falling to $84 million in 2025, the lowest on record.
For users, the defense is narrow but effective: never compile and deploy code from a tutorial's chosen website, verify the deployed bytecode against the source you pasted, and treat any "bot" that needs your deposit more than your skills as the scam it almost always is. For platforms, the harder question is whether YouTube and ad networks can be pressured to catch AI-generated tutorial farms before they reach 224 victims. For US victims, the FBI's Internet Crime Complaint Center accepts crypto fraud reports, and the bureau logged $16.6 billion in reported internet crime losses in 2024. The next version of this scam will use a better host, a bigger name in AI, and a more convincing compiler. The structure, a victim building his own trap, will not change.












































