Fake AI trading-bot tutorials gave scammers a quieter route than a hostile wallet pop-up: they persuaded victims to deploy and fund the drainer themselves. The useful warning is not that Claude or any AI model hacked a wallet. It is that the promise of an AI-built arbitrage bot made a dangerous developer workflow look routine.
| What you need to know |
|---|
| 1TRM Labs says nine similar YouTube tutorials pushed viewers toward fake AI trading bots. |
| 2The firm traced 274.60 ETH from 224 victim wallets into six shared collection addresses between 12 February and 11 August 2026. |
| 3The attack did not rely on a seed-phrase form, spoofed exchange login or malicious token approval. |
| 4The narrow lesson is to treat a browser-based compiler or development tool as part of the security boundary. |
In a 14 September report, TRM Labs said the videos promised a Claude-assisted arbitrage bot and then walked viewers through wallet setup, code copying, contract deployment and funding. TRM traced 274.60 ETH, worth about $517,205 at the time of the transfers, from 224 victim wallets into six shared collection addresses. The observed drains ran from 12 February to 11 August 2026.
That loss total is not large enough to make the case a market event. The method is the story. The victim signs the deployment and funding transactions from their own wallet, so the usual mental model of a phishing site asking for a seed phrase or a suspicious approval does not fit. The contract is fresh, the wallet action can look intentional, and the theft is hidden inside the workflow the tutorial just taught.
The compiler became the attack surface
TRM says the operation used a fake compiler styled to look like a normal smart-contract development tool. In one variant, a background script discarded the source code pasted by the victim and compiled a different contract fetched from the attacker’s server. That matters because a readable source-code window or a friendly comment section does not prove which bytecode reaches the chain.
The deployed contract had no trading logic, according to TRM. It forwarded funds to operator-controlled addresses when the victim called functions such as Start or Withdraw. TRM also described a fake “gas nonce liquidity” message that encouraged some users to add more ETH before the drain, with small balances below 0.05 ETH helping the scheme look less alarming during early testing.
AI was bait, not the bot
The Claude branding should be read narrowly. TRM says it found no Anthropic product and no AI function inside the deployed contracts or compiler sites. The AI name made the pitch more believable because no-code trading bots and AI coding assistants are now familiar enough to feel plausible. TRM also says earlier versions of the same playbook used ChatGPT as the lure, which points to rotating branding rather than an AI-specific exploit.
That caveat also keeps the reader from learning the wrong lesson. Blocking a model name would not solve this. The failure point is trust in a tutorial-supplied development environment, especially when the user does not independently verify the compiler, deployed bytecode, contract address and on-chain behavior before adding funds.
Why normal warnings can miss it
Wallet prompts can show that a deployment or transfer is happening, but they cannot explain the off-chain tutorial context. A user who believes they are launching their own arbitrage bot may approve the exact transaction the attacker needs. That makes this pattern different from a fake exchange login, a malicious approval prompt or a contract interaction with a known scam address.
TRM says the funds later moved through DeFi, bridges and a mixer, with no centralized-exchange hop in the outbound flow it observed. That reduces the easy recovery path and makes early prevention more important. US victims can still file crypto-fraud reports through the FBI’s IC3 cryptocurrency complaint page, but the better control is earlier: do not let a tutorial choose the compiler, and do not treat visible source code as proof of deployed behavior.
