How Harvey secures embeddings at scale
Protecting sensitive data from embedding reversal through secure-by-design architecture.
Helping legal and professional service teams leverage AI effectively and securely means that we deal with some of the most sensitive data in the world. Much of this sensitive data powers our enterprise-grade RAG systems.
Previously, we discussed how we build these systems at scale by choosing a vector database that prioritizes security, performance, and reliability and establishing strong management of customer data. Embeddings are the foundation of our enterprise-grade RAG systems, but they come with an emerging security threat: embedding reversal.
At Harvey, we’re building a platform that is secure by design and stays secure as our product and the industry evolve. In this post, we describe how we’ve applied that approach to embedding reversal: what recent research has shown is possible, why it matters for the sensitive work our customers bring to us, and the architectural controls we’ve designed as a result.
To read the article in full, click here.
