uri-eliabayev talks with shiran-bareli, VP Research at Cyera (founded 2021; data security posture management, DLP, and now AI security). Classifying sensitive data across hundreds of millions of files a day rules out frontier LLMs on cost. Cyera splits the task into entity-level detection (credit cards, names) and document-level sensitivity (an M&A memo is sensitive even with no personal data), uses big LLMs plus a cross-functional committee and judges to build ground truth, then trains its own encoder-decoder models, self-hosted and GPU-optimized. The result is cheaper, faster, and more accurate, because context matters (a lawyer’s brochure phone number isn’t sensitive). AI assistants plugged into Drive, M365, and Notion made the old access-governance problem far harder: over-permissioned users can now simply ask. Agents add intent problems: inventory every agent, check what data it can reach, detect out-of-scope actions, and watch for injected or insider prompts. Also: match architecture to the problem (encoders for extraction, fine-tuned open decoders for explained risk scoring); a dedicated team drives safe AI adoption for non-technical staff; vulnerability research is their biggest token user (“a giant magnet for the needle in the haystack”).
Show: hidden-layers
Concepts: small-language-models · agent-security · model-selection · llm-evals