AI Agent Loses Up to $200K in Crypto After Morse Code Prompt Injection Attack

An attacker drained between $150,000 and $200,000 in cryptocurrency from a wallet controlled by xAI's Grok-connected trading bot Bankr by posting obfuscated instructions on X. The malicious content, disguised as Morse code strings and fragmented Python snippets, bypassed standard keyword filters because they do not decode encoded text before scanning. Grok's AI agent ingested the social media post as regular content but interpreted the decoded payload as a legitimate transfer command, which Bankr then executed without verifying the instruction's source. No traditional software exploit was involved; the attack succeeded solely because the agent lacked an architectural boundary separating untrusted external content from authorized operator commands. Security researchers classify this as OWASP's top agentic AI risk, ASI01 Agent Goal Hijack, highlighting a systemic vulnerability in AI agents that hold execution privileges over financial tools.