Hi there,

AI systems consume a lot of data and the risks around these systems are very real. I spoke with a group of privacy professionals during a webinar on "How to address privacy needs across the AI tech stack." (If you'd prefer to scroll through the slides, you can view the PDF of the presentation here.)

Most of the security conversations happening right now concern the AI model. But there are actually three main areas of modern AI systems where you need to evaluate risk: the AI model, the AI memory, and the interactions.

The AI memory is a newer aspect of AI systems, driven by the need to reduce hallucinations and provide context, and as a way to build anomaly detection, recommendation, classification, and other systems with up-to-date information that doesn't require a custom model.

Securing the AI memory is something I've written about a lot recently because it's one of the biggest security gaps. If you're not familiar with vector databases or vector embeddings, I recommend reading my blog post on embedding myths.

For a comprehensive look at how data flows within AI systems and where it is stored (and vulnerable), you can watch the full webinar on-demand on YouTube. 

Protect the AI Memory 

Data stored in the AI memory needs to be protected. It will take the form of either vector embeddings or metadata stored in a vector database (there are many to choose from right now). None of these databases offer you encryption-in-use. But you can still encrypt and use your data with our newest product, Cloaked AI, which launches out of beta in two weeks. You can access the beta now.

Thanks for reading along. Until next month,

Patrick Walsh
CEO, IronCore