Threat Modeling For LLMs

LLM Red Teaming: Building Defenses Against Adversarial AI Attacks

Your model passed standard safety benchmarks. A simple vendor email prompt just exfiltrated internal pricing...