OpenAI Releases PII Filter

OpenAI has released Privacy Filter. It is a bidirectional token-classification model that finds and masks PII in text, designed for high-speed data cleaning workflows.

• It is provided under the Apache 2.0 license, enabling experimentation, customization, and commercial deployment.
• It has 1.5B parameters and 50M active parameters, allowing it to run in a browser or on a laptop.
• It supports a 128,000 token context, enabling long documents to be processed without chunking.
• openai/privacy-filter is based on a structure similar to gpt-oss, and was subsequently converted into a classifier and post-trained to fit a privacy label taxonomy.

The model labels the input in a single pass, then reconstructs BIOES spans using constrained Viterbi decoding. The output labels consist of 8 categories: account_number, private_address, private_email, private_person, private_phone, private_url, private_date, and secret, and it predicts 33 classes at the token level.

At runtime, operating points can be configured to adjust precision/recall and masking span length. However, OpenAI has defined this model as a redaction aid, not an anonymization guarantee tool, and noted that performance degradation and false positives/negatives may occur with non-English text, rare names, and domain-specific identifiers.