Although the prompt injection vulnerability in a certain exchange's AgentKit has been addressed, its actual impact has been significantly underestimated.

According to a report from CriptoNoticias, an independent security researcher has disclosed a prompt injection vulnerability in a certain exchange's AgentKit, which allows attackers to induce the AI agent to execute unauthorized token transfers via malicious instructions, without requiring manual confirmation.

This vulnerability has been validated through actual transactions on the Base Sepolia testnet. Additionally, the researcher pointed out that the vulnerability also exposes the unlimited authorization process for ERC-20 tokens, as well as remote server access permissions within the same execution context of the agent, expanding the risk beyond merely having the wallet drained. However, the report did not specify which particular infrastructures might be affected.

The vulnerability was submitted to the exchange’s bug bounty program in February and was officially validated. It was ultimately classified as medium risk and awarded a $2,000 bounty. However, the researcher emphasized that the actual impact of the vulnerability is far greater than the official rating.