I've briefly touched it in another blog post. It is not just trusting the institutions to not use the data for training. Optimizations you apply to making LLM i

I've briefly touched it in another blog post. It is not just trusting the institutions to not use the data for training. Optimizations you apply to making LLM inference fast might present new security landscape that we don't know much about. For example, if your LLM inference engine stores KV cache across user sessions, you can in theory do a timing attack to retrieve what's in the KV cache, potentially leaking other users' requests.