Adversarial ML: Extortion via LLM Manipulation Tactics · Blog · Sublime Security

This report analyzes an in-the-wild extortion email campaign targeting Google Workspace recipients that leverages prompt-injection phrases like "IGNORE EVERYTHING ELSE" and sender spoofing to try to bypass LLM-based phishing detectors and pressure victims into cryptocurrency payments; Sublime identified the attempt using extortion- and crypto-related signals and notes its BERT-based NLU does not follow in-text instructions.