Data security startup Immuta Inc. today announced new data governance and audit capabilities for retrieval-augmented generation-based generative artificial intelligence solutions across multiple cloud platforms.

Retrieval-augmented generation applications supply large language models with external knowledge sources to improve the accuracy and relevance of generated content. RAG-based models also retrieve pertinent information from external databases or documents, which is used to inform the generation process.

With today’s release, Immuta is offering multilayer architecture for securing, monitoring and auditing sensitive data accessed by RAG-based AI applications. Doing so allows those in charge of overseeing data governance to de-risk their data and take control of generative AI security at the data layer.

The issue the release seeks to tackle is an increasingly common one – the difference between corporate data management concerns and employee use of AI. In a recent survey, Immuta found that 80% of data experts believed that AI is making data security more challenging, while at the same time, 88% say their employees are also using AI irrespective of whether the company has officially adopted it. This causes friction between AI users and information technology workers as well as introduces the risk of rogue or unsanctioned AI tools.

Immuta argues that there are three lines of defense when it comes to securing AI applications — the storage layer, the data layer and the prompt layer. “While the initial focus for early adopters securing AI applications was at the prompt layer, this should not be the only focus – in fact, it should be treated as the last line of defense,” said Chief Product Officer Mo Plassing. “With this release, we’re focusing on those first two lines of defense: the storage and data layers.”

The company has collaborated with Amazon Web Services Inc. to develop a native Amazon S3 storage integration that enforces fine-grained and scalable access control on unstructured data stored in S3, to address the first layer of AI application security — the storage layer. Through the feature, attribute-based access controls are pushed down to the storage layer to secure the first line of defense.

The data layer, the second line of defense, is where unstructured data is transformed for model training and “chunked” for Retrieval Augmented Generation use cases. With today’s update, RAG indexes can now be discovered, classified and controlled in the same manner as other traditional data sources.

With the new solutions, data teams can control access to the storage layer with multilayered policies for securing sensitive data when building RAG indexes. That way, they can maintain a highly accurate and granular metadata inventory of RAG indexes with topic-based classification and control access to RAG-based applications through the data layer. This gives data platform teams control through natural language policy creation, prompt/query-time policy enforcement, multiplatform RAG support from Snowflake and Databricks, and domain-specific RAG policy.

In addition, users can monitor and audit RAG index access with operational monitors, which provide a continuous view of RAG operations and a single view of AI application data access across all supported platforms.

“Data teams are now able to leverage the significant investments they have made in their cloud data platforms and rapidly extend this work to their AI application workloads,” Plassing added. “With Immuta, they now have a single control plane for policy enforcement, visibility and auditing that works across multiple cloud platforms and RAG models.”

Photo: Immuta