OpenAI published a token-classification model on Hugging Face for detecting and masking personally identifiable information in text. The model supports high-throughput on-premises sanitization with a 128000-token context window and runs with 1.5B total parameters and 50M active parameters (https://huggingface.co/openai/privacy-filter). It was pretrained autoregressively on a gpt-oss checkpoint then converted to a bidirectional classifier using supervised token-level classification. The model predicts over an 8-category privacy taxonomy and applies constrained Viterbi decoding to produce coherent BIOES spans. OpenAI released it under Apache 2.0. The EU AI Act classifies systems processing personal data as high-risk and mandates appropriate technical safeguards (https://artificialintelligenceact.eu/). A 2023 arXiv survey on LLM privacy attacks documented training-data extraction risks that persist across production pipelines (https://arxiv.org/abs/2310.10078). Coverage of the release described architecture details but omitted explicit linkage to the regulatory compliance timeline and the shift toward local data fabric tooling now required by enterprise deployment standards.