The Security Leader's Guide to IronCore Labs

Once you encrypt at the application layer, a long list of risks collapses at once: a database admin can no longer browse customer data; a leaked read-replica or backup is useless; a subpoena to your cloud provider returns ciphertext; with per-tenant keys, an injection attack can't exfiltrate across tenants. Multi-tenant SaaS gets the trust posture of single-tenant on-prem deployments without the cost.

The attacker needs to gain access to the key and the data or the already decrypted data. Push it one step further with customer-held keys (HYOK) and you remove yourself and your employees from the trust boundary entirely, making an attacker's job exponentially more difficult.

Using ALE to restrict insider access →
Avoiding compelled access with encryption →
How SaaS handles sensitive data →

AI needs data; not just for training, but at inference time. Without auxiliary data, models are more prone to hallucinate. With supporting data, a model can cite its source. So we jam context like the contents of web pages or documents into prompts invisibly as part of interactions with AI. This is done in a process called RAG (Retrival-Augmented Generation), and it’s the path that introduces private data (company documents, our emails, calendar, SaaS data, etc.) into AI interactions.

Vector embeddings underpin much of the AI search and generation process. These embeddings are mathematical representations of data that can be used by AI, and especially RAG workflows, to find information relevant to a particular query. The vectors are typically generated by purpose-built AI models called embedding models, and they are typically stored in vector databases or vector-capable databases.

Vector database advocates have claimed that these vectors are unimportant from a security or privacy perspective, because they distill the information from the source data in a one-way process (like a hash). This is dead wrong.

Vector embeddings used in RAG and semantic search are not irreversible. They can be inverted back to near-perfect approximations of the original text, faces, or images. Every vector in your architecture is, effectively, another copy of your private data with weaker access controls, lowered visibility, and almost no protection.

IronCore’s Cloaked AI is the practical answer: encrypt vectors, keep them searchable, and stop creating unprotected shadow copies of your sensitive data.

RAG security risks explained →
There and back again: an embedding attack journey →
Talking to the business about AI risk →

Securing Gen AI white paper →