Adversa AI researcher found an injection vulnerability in Grok and Gemini

Adversa AI described a Cryptographic Context Injection technique that uses AES-encrypted instructions to bypass AI safety filters. In the Grok demonstration, hidden instructions from a webpage caused the model to process private session data and the user’s full chat history and send it to an attacker-controlled website without a user click or warning. In the Gemini demonstration, decrypted output influenced the model’s behavior and caused it to generate content that the model would normally refuse.