Encrypted Prompts Defeat Grok and Gemini Guardrails; Chat Histories Stolen

Adversa AI disclosed "Cryptographic Context Injection," a technique that embeds AES-256-GCM ciphertext in web content so agentic LLMs decrypt and execute hidden instructions inside their code-execution sandboxes, bypassing text-based guardrails; researchers showed zero-click exfiltration of Grok chat histories and dangerous output from Gemini, highlighted limited vendor engagement and disclosure gaps, and warned that content-based guardrails cannot inspect payloads that only exist in plaintext after model execution.