Adversa AI Shows How Encrypted Prompts Bypass Grok and Gemini Safety Filters

Cybersecurity firm Adversa AI demonstrated a zero-click attack in which AES-encrypted malicious instructions embedded on a webpage bypassed the text-based safety filters of both Grok and Gemini AI assistants. Because the payload was ciphertext, no content filter flagged it as harmful during the input stage. When the AI models used their built-in code execution capabilities to decrypt the blob, the resulting plaintext was treated as trusted internal output rather than external untrusted content. This provenance loss allowed the decrypted instructions to direct the model to make an outbound request, exfiltrating the user's chat history, name, and location to an attacker-controlled server without any user interaction. The vulnerability highlights a fundamental trust-boundary flaw in AI agents that combine code execution with unsanitised web content, which standard pattern-matching guardrails cannot address.