Practical LLM Security Advice from the NVIDIA AI Red Team | NVIDIA Technical Blog - NVIDIA Developer

The NVIDIA AI Red Team highlights critical vulnerabilities in LLM-based applications, most notably Remote Code Execution (RCE) via prompt injection when LLM-generated code is executed using functions like `exec` or `eval` without sufficient sandboxing. Other significant findings include data exfiltration risks from active content rendering in LLM outputs and data leakage through insecure access controls in Retrieval-Augmented Generation (RAG) data sources.