Encrypted Prompts Bypass AI Safety Guardrails in Grok and Gemini

Researchers have discovered a technique called Cryptographic Context Injection, which allows malicious instructions to be concealed until they are decrypted within a trusted environment. This method enables bypassing of safety guardrails in AI systems, specifically in Grok and Gemini. The technique involves encrypting prompts that are then decrypted inside the trusted execution environment, effectively hiding the malicious instructions until they are executed.