I had the pleasure of introducing my co-founder, Abhishek Das, at DBTA’s annual Data Summit in Boston. Abhishek was invited to present one of the keynote sessions, and we thought it would be a good opportunity to discuss the emerging security challenges as organizations look to harness their proprietary data as part of their artificial intelligence (AI) strategy.
AI has become a C-level mandate, and every organization is looking to gain a competitive advantage. However, despite this incredible momentum, less than twenty percent of AI projects have reached the production stage. Security of AI applications, and specifically confidentiality and privacy concerns about the data we feed into them are a major reason for these roadblocks.
AI teams looking to leverage their proprietary data and eliminate hallucination issues, go beyond basic prompt engineering approaches to use Retrieval Augmented Generation (RAG) and Fine-tuning architectures. In fact, 75% of enterprises have already started using these approaches. These architectures were the focus of Abhishek’s presentation, and specifically the data security challenges that come to the fore here. Abhishek, who has decades of experience working with AI and machine learning (ML) applications, outlined a 3-layer stack of such AI applications. Starting from the Inference Layer with the user/application interfaces, to the Model Layer and finally the Data Layer at the bottom.
Most AI security efforts to date have focused on the Inference layer analyzing prompts/responses, and some on the Model Layer looking at model supply chain issues. However, it is at the Data Layer, where new and critical security and privacy concerns need to be addressed to successfully productionalize RAG & fine-tuning architectures.
Abhishek detailed 6 primary data security risks. These are (i) data privacy, (ii) training data poisoning, (iii) prompt manipulation, (iv) unauthorized access, (v) sensitive data exfiltration and (vi) data supply-chain poisoning. These risks can be clearly mapped to industry standard AI security frameworks such as the OWASP LLM Top 10 and the Databricks AI Security Framework.
He wrapped up the talk with a brief glimpse of how Acante is squarely addressing these security risks at the Data Layer of the AI stack. Every single bit of organizational data will be at risk of exposure through these AI systems in ways we never fathomed before and ways that aren’t easily comprehensible to humans. Talk to us about how Acante is empowering enterprises with practical ways to unlock the value of their data safely and confidently.
Watch the whole keynote above or checkout the slides on our LinkedIn page.
