When AI's Helpfulness Becomes a Liability: The Morse Code Attack on Grok

This attack exposes a fundamental architectural vulnerability in AI agent systems: when AI processes untrusted external content and passes decoded output directly to downstream execution systems without validation, attackers can inject commands through any encoding scheme. By disguising a transaction request in Morse code, the attacker bypassed keyword filters and exploited Grok's helpful nature to steal $200,000—a pattern that threatens any AI system handling untrusted input.