Adversa AI, an AI security company, disclosed a novel attack technique called Cryptographic Context Injection that can exfiltrate private data from AI chatbots like Grok. The attack uses encrypted payloads to bypass content filters and execute malicious instructions in the model's runtime.