Grok wallet drained via prompt injection using Morse code obfuscation
An attacker used a Morse code-obfuscated prompt injection to trick Grok, an AI system managing a cryptocurrency wallet, into transferring 3 billion DRB tokens worth approximately $155,000 to $200,000 to an unauthorized wallet. Most funds were reportedly returned following public pressure, but the incident exposed vulnerabilities in how AI agents with execution permissions validate user instructions.
What happened
On 4 May, an attacker exploited a Grok-linked cryptocurrency wallet by sending a Bankr Club Membership NFT to the wallet, which unlocked advanced tool permissions within the Bankr system.
The attacker then crafted a malicious prompt using Morse code obfuscation and social engineering techniques. Grok decoded the Morse code into a clean public instruction, and Bankrbot executed it as a valid command, resulting in an unauthorized transfer of 3 billion DRB tokens to an attacker-controlled wallet.
The transferred tokens were valued at approximately $155,000 to $200,000 at the time and were subsequently transferred again and rapidly sold.
Reported impact
The attack drained the Grok-linked wallet of approximately $155,000 to $200,000 in cryptocurrency.
Following public pressure, reports indicate that 80% to 88% of the funds were returned in ETH and USDC, with the remaining amount potentially retained as an informal bug bounty. Recovery details were not fully verified through official statements at the time of reporting.
Index analysis
The incident did not exploit a vulnerability in smart contracts or blockchain infrastructure. Instead, it targeted intent parsing, tool permission systems, and AI decision-making layers by manipulating how Grok interpreted user input.
The attack highlighted a vulnerability in multi-agent systems where one AI system decodes user instructions and another system with execution permissions treats that output as valid. Wallet permissions, parser, social trigger, and execution policy became layers of attack vectors.
Governance controls
The exposure occurred because user input was not properly constrained before execution. The incident demonstrates how AI agents with execution capabilities can become vulnerable when given sufficient tool permissions without adequate safeguards on instruction validation.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.