IronCore Labs Cloaked AI protects vector embeddings

IronCore Labs launched Cloaked AI, an SDK that protects vector embeddings with data-in-use encryption.

Large language models are shifting the paradigm for how AI products are built and where private data is stored. While private AI data used to be in the model, now it’s captured as embeddings and stored in vector databases. These databases store everything from internal proprietary documents to chat histories and medical diagnoses.

“We’re thrilled to launch Cloaked AI and provide a solution that allows companies to protect sensitive AI data while also maintaining full functionality,” said IronCore Labs CEO, Patrick Walsh. “Cloaked AI keeps this new class of data secure and usable with strong encryption, unbypassable access controls, and optional bring your own keys (BYOK) and sovereign data functionality. And companies can still use whichever vector database best supports their use case.”

Key features of Cloaked AI:

• Built for easy integration and quick adoption with just a few lines of code
• No new services or containers required
• Integrates with existing vector database or FAISS indices
• Access to data is restricted to permissioned systems with no backdoor access by hosting providers or database admins
• Easy to add on advanced key management and BYOK functionality with IronCore Labs SaaS Shield
• Prevents embedding inversion and membership inference attacks
• Protects associated sensitive data/metadata as well as vectors
• Works with the most popular generative AI use cases including RAG
• Highly performant with negligible impact on query latency

Cloaked AI is an encryption-in-use solution that protects vector embeddings without compromising usability or hampering AI use cases like anomaly detection, biometric identification, semantic search, and so on. Cloaked AI works with all known vector databases, including those from Pinecone, Weaviate, Qdrant, Elastic, and AWS OpenSearch.