Prompt Injection via Tokens: The Dark Side of Agentic Commerce

Evidence record

Blockaid threat-intelligence post describing malicious instructions embedded in token metadata and a staged exchange-agent demonstration in which a token description induced an unlimited approval that the agent's signer executed. The page does not establish a production loss from that staged vector.

Read the original source ↗

Claim-level citations (3)

• supportsToken metadata prompt injection reached an on-chain agent's signer in a staged drain: In the staged September 14 exchange-agent demonstration described by Blockaid, a planted token description was interpreted as an instruction to grant an unlimited approval, and the agent's signer executed it.

  Section 'Staged Drain - Exchange Agent (Demo, September 14)'

• supportsToken metadata prompt injection reached an on-chain agent's signer in a staged drain: Blockaid reported observing malicious prompt-like instructions embedded in live token metadata fields such as names, symbols and descriptions.

  Opening section describing prompt injections embedded in token names, symbols and descriptions

• supportsToken metadata prompt injection reached an on-chain agent's signer in a staged drain: The Blockaid disclosure labels the unlimited-approval path as a staged demonstration and does not establish a production victim loss caused by that exact token-metadata vector.

  Section 'Staged Drain - Exchange Agent (Demo, September 14)'