0xGem.base.eth

What's the trust boundary between an AI agent and an onchain execution layer? The BankrBot/Grok incident is a case study on this very question BankrBot is an AI agent designed to let people execute crypto transactions through X and Farcaster In an early experiment for AI-to-AI interaction, Grok was asked to come up with a name for a token and proposed DebtReliefBot $DRB. BankrBot automatically deployed the token to Base, setting up Grok with a wallet that receives some fees from DRB activity On May 4th, an attacker sent Grok a Morse-code prompt injection and asked for it to translate the message, essentially making Grok relay a command to BankrBot BankrBot treated it as an executable instruction and transferred approximately 3 billion DRB worth ~$175,000 from Grok’s wallet to the attacker's wallet The smart contracts worked as designed, and the message bypassed content safety checks While ~80% of the funds were ultimately returned due to community pressure (the remaining ~20% held as a bug bounty), it still raised a structural security concern AI agents with onchain execution rights are becoming a popular part of DeFi infrastructure. The security standards around them need to keep pace If you’re working at the intersection of AI + onchain and want to keep your digital assets secure, get in touch with our team You’re Covered with Nexus Mutual 🤝