Encrypted Prompts Bypass AI Safety Guardrails in Grok and Gemini

Researchers say the new ‘Cryptographic Context Injection’ technique conceals malicious instructions until they are decrypted inside a trusted execution environment. The post Encrypted Prompts Bypass AI Safety Guardrails in Grok and Gemini appeared first on SecurityWeek .

The article reports on a technique called cryptographic context injection where attackers encrypt malicious instructions that are only decrypted and interpreted inside the model’s trusted execution environment, allowing them to bypass safety guardrails in systems like Grok and Gemini. This is a documented method of hiding harmful prompts so they appear benign to external safety filters but execute as malicious instructions once processed by the model. From a RealGround perspective, this illustrates that safety checks operating only on visible prompt text are insufficient; controls must account for encrypted or obfuscated instructions and where decryption occurs in the AI pipeline. Practically, organizations should harden agent architectures against hidden prompt injection, implement deeper red-teaming for encrypted/encoded inputs, and enforce policies on how and where cryptographic operations interact with model context.