MPIB: A Benchmark for Medical Prompt Injection Attacks and Clinical Safety in LLMs
The episode reviews "MPIB: A Benchmark for Medical Prompt Injection Attacks and Clinical Safety in LLMs." This discussion a massive dataset of over 9,697 curated instances used to test real-world threats like RAG-mediated injection. The hosts conclude that surface compliance is insufficient for patient safety, emphasizing the need to build trustworthy AI systems based on measurable clinical outcomes.