New Cryptographic Context Injection Attack Bypasses AI Safety Filters to Steal Chat Histories

Adversa AI researcher Rony Utevsky developed a new attack technique called Cryptographic Context Injection, which bypasses AI safety filters by embedding instructions in AES-encrypted ciphertext. The method tricks AI models, including Grok, into decrypting the payload within their runtime, enabling zero-click theft of full chat histories. The attack exploits the model’s ability to process encrypted data without requiring user interaction or explicit consent. No specific dates, CVE IDs, or affected version ranges were disclosed in the report. The impact includes unauthorized access to sensitive conversational data stored by AI systems.