The important shift is architectural: teams can mask sensitive text before it ever leaves the machine. OpenAI’s 1.5B-parameter Privacy Filter supports 128,000 tokens and scored 97.43% F1 on a corrected version of the PII-Masking-300k benchmark.
Databricks posted on March 27, 2026 that its LogSentinel system uses LLMs to classify columns, apply hierarchical and residency-aware labels, and detect drift, with up to 92% precision and 95% recall for PII on 2,258 samples. Databricks documentation says Unity Catalog Data Classification uses an AI agent and LLM to classify and tag tables, while governed tags and ABAC policies translate those tags into consistent access and compliance controls.