Grok AI vulnerable to encrypted prompt attacks that leak user data

Security researchers have discovered a new attack method called Cryptographic Context Injection that can bypass safety guardrails in Grok, xAI's large language model. The technique involves embedding malicious instructions within encrypted content, which the model processes without recognizing as a threat. This allows attackers to manipulate Grok into exfiltrating sensitive user data during a conversation. The vulnerability highlights a broader pattern of adversarial methods being used to circumvent LLM safety mechanisms. It represents one of the latest in a growing series of techniques that undermine built-in protections in AI systems.