Disclosed are various approaches for large language model (LLM) interaction security sandboxing. A client device can execute an LLM security sandbox that includes at least one LLM communications sanitization process. The LLM security sandbox can perform the at least one LLM communications sanitization process on the LLM message to generate an approved LLM message. The client device can provide access to the approved LLM message by at least generating a user interface that includes the approved LLM message, or transmitting the approved LLM message from the client device to the LLM service.