an X user just tricked Grok into sending them $200k in crypto by hiding the instruction in morse code. new exploit category??
so this happened a couple days ago and i can't stop thinking about it.
an X user sent a Bankr Club Membership NFT to Grok's wallet, which expanded Grok's permissions inside the Bankr trading bot system. then they prompted Grok to translate a morse code message and pass it to Bankrbot. the decoded message was a transfer instruction. Bankrbot processed it as a valid command. 3 billion DRB tokens (~$200k) sent to the attacker's wallet on Base. attacker dumped immediately, deleted the X account, walked.
morse code. this is where we are now.
the lesson isn't that Grok is dumb. the leson is we've started giving AI agents wallet permissions and the attack surface is enormous. an AI with wallet access. permissions that expand via NFT transfers. trust relationships between AI systems. translation features that don't sanitize output. each is a normal feature in isolation. combined, it's a disaster.
every additional layer of integration adds attack surface. AI agents are a maximum-integration play. they read prompts, parse contexts, hold credentials, talk to other systems, execute on-chain. each interface is a potential injection point.
what protects you from this while swapping. simple immutable contracts.
take Sushi's AMM pools as the textbook example. immutable code. no AI in the loop. no permission system to expand. no translation feature to abuse. you swap, the math executes, done. you can't social-engineer a smart contract that has no admin functions, because there's nothing to talk to.
we keep finding out the hard way that LLMs can be tricked. prompt injection has been a known issue for years. now we're stapling wallets to LLMs and expecting it to be safe.
how worried should we be about the broader AI-agent-wallet pattern?