Grok Vulnerable to Data Exfiltration via Cryptographic Context Injection

Researchers have discovered a vulnerability called Cryptographic Context Injection that allows attackers to bypass safety guardrails on xAI's Grok. By encrypting malicious instructions, attackers can trick the LLM into exfiltrating sensitive user data. This vulnerability highlights a critical flaw in current LLM safety guardrails, which often fail to inspect or understand encrypted inputs. It demonstrates that traditional input-filtering defenses are insufficient against sophisticated context manipulation attacks. The attack bypasses standard prompt-monitoring systems because the malicious payload remains obfuscated until processed within the model's runtime context. Once executed, the model can be coerced into exfiltrating data through channels like markdown images or external API calls.