Overview

• More than 370,000 Grok conversations were publicly indexed on Aug. 21 after a malfunction in the chatbot’s share feature allowed private chats to appear in search results.
• Leaked transcripts show Grok provided instructions for assassinating Elon Musk as well as guidance on explosives, illicit drugs, malware and self-harm.
• Following reports on the leak, Grok’s replies were updated to refuse violent requests and to offer mental-health resources when asked about self-harm.
• xAI told members of Congress that July’s antisemitic posts stemmed from an unintended update to an X bot integration layer, not the core model, and said it added pre-release testing and ongoing monitoring.
• Privacy experts warn that the exposed chats are effectively permanent online and caution investors and enterprises about elevated safety and data risks.