Blockchain analysis firm TRM Labs has revealed that a series of YouTube tutorials claiming to teach viewers how to build AI arbitrage bots using "Claude" are actually scams that lure viewers into deploying malicious smart contracts. TRM tracked that between February 12 and August 11, 2026, these tutorials led to the theft of 274.60 Ethereum (ETH, approximately $517,205 at the time of theft) from 224 wallet addresses.

The scam process involves the tutorial first enticing viewers to create a cryptocurrency wallet step by step, then guiding them to copy and paste "code" into a counterfeit compilation website that mimics the official Ethereum development tool, Remix. The counterfeit site quietly discards the clean code pasted by viewers on the server side, instead compiling and deploying malicious bytecode provided by the attacker's server, while viewers see "clean" source code on their screens. The tutorial then instructs viewers to inject 1 to 2 ETH into the newly deployed contract to "activate arbitrage functionality"; once the account balance exceeds 0.05 ETH, clicking the "start" or "withdraw" button in the tutorial will automatically transfer funds to the attacker's address.

Since the entire process is completed by the victims using their own wallets, without involving common phishing links, counterfeit domains, or malicious authorization signatures, it does not trigger the usual security warnings or authorization analysis alerts from wallets. This is a key reason why such scams can bypass mainstream wallet protection mechanisms. Some victims, after their initial funds were stolen, received a false error message claiming they needed to pay a "gas nonce liquidity" fee (a term that does not exist in the Ethereum technical framework), luring them into a second round of transfers.

TRM Labs identified 9 YouTube tutorials with nearly identical content, seemingly from different creators, but all using AI-generated host images and voiceovers, with completely consistent scripts. Even specific profit promises like "earn 1 ETH every 20 hours" are identical to the decimal point; these videos also fake user reviews in the comments section and hide key links in the video description to evade scrutiny. As of September 2026, these 9 videos have accumulated 310,474 views, with earlier versions having been removed by YouTube.

The stolen 274.60 ETH involved 224 victim wallets and 234 deployed malicious contracts, with a median loss of about 1 ETH per victim. TRM found that these funds were entirely transferred through decentralized channels—exchanging ETH for Dai stablecoin, transferring via cross-chain bridges, and laundering through mixing protocols, without passing through any centralized exchanges, ultimately splitting the funds into six addresses controlled by the attackers.

TRM Labs also released the "2026 AI Crime Application Index" report, showing that criminals (led by scammers) have increased their use of AI technology by 40% year-on-year. The methods of disguising scam videos as educational content using AI-generated material and mass copying of scam templates are typical manifestations of this growth trend. Notably, TRM clarified in the report that Claude itself was not actually involved in the writing or operation of these malicious contracts—no traces of any Anthropic products or AI functions were found in the deployed code, and the name "Claude" was merely used as a marketing gimmick for the scam. Similar scams using almost identical methods had previously appeared in 2025, using the name ChatGPT, indicating that attackers can switch popular AI brand names at any time to package the same scam script, with the real damage being to the reputation of the impersonated AI brand and inexperienced investors, while the attackers bear almost no technical development costs, only needing to change the brand name and packaging materials to cash out repeatedly.

Reports indicate that earlier versions of the tutorials used the name ChatGPT, employing the exact same scam mechanism, demonstrating that this "AI arbitrage bot" scam template is highly replicable, with attackers merely changing the marketing facade based on the most attention-grabbing AI brands at the time.

Source: Public Information