Encrypted Prompts Bypass AI Safety Guardrails in Grok and Gemini

Researchers at Adversa AI discovered a new attack technique called ‘Cryptographic Context Injection’ that bypasses AI safety guardrails in xAI’s Grok and Gemini models. They disclosed their findings to xAI in June 2026, but have not received a response. The attack involves delivering encrypted prompts to the AI models, which then execute the instructions within the model’s code execution sandbox, bypassing safety filters. The attack can be delivered directly or through a watering hole attack. Despite the decreasing success rate against Gemini, the potential for this attack remains a significant concern.