Blockaid warned that attackers can hide prompt-injection instructions in token names, symbols and descriptions, creating an additional attack surface for AI agents that analyze tokens or execute blockchain transactions. The risk increases when an agent can access a crypto wallet or sign transactions. Blockaid said Solana recorded more than 263,000 new tokens in a single day this month, with pump.fun responsible for most launches, increasing the volume of metadata that automated systems may process. The firm cited a May 4 incident in which an AI-controlled wallet associated with Bankr transferred approximately $215,000 in tokens after being prompted publicly through social media, a September 3 incident in which a user lost approximately $2.1 million after ChatGPT directed the user to sceptre.network to swap a token, and another reported loss of about $500 through a similar website. A September 14 demonstration showed how manipulated metadata could potentially instruct an exchange AI agent to approve an unlimited token allowance, although Blockaid described it as staged rather than a reported theft. Traditional token checks may not detect malicious instructions in metadata because they focus on smart-contract behavior, liquidity and links. Transaction simulation can add a checkpoint by showing expected asset movements, permissions and contract calls before signing, while Blockaid said MetaMask has routed Agent Wallet transactions through its validation since June and introduced AI scam detection for its broader user base in September. The incidents use different attack methods but highlight the security boundary between information an agent can read and instructions it is authorized to follow.