OpenAI releases Privacy Filter as an open model for PII redaction

OpenAI has launched Privacy Filter, an open model for detecting and masking personally identifiable information in text. The company frames it as infrastructure for more resilient, privacy-preserving AI systems, with a strong emphasis on local execution so raw data does not need to leave the machine.

What is new

Privacy Filter is a small token-classification model with 1.5 billion total parameters and 50 million active parameters. It supports up to 128,000 tokens of context and can label eight types of sensitive information, including names, addresses, emails, phone numbers, account numbers, and secrets such as API keys and passwords. OpenAI reports a 96% F1 score on the PII-Masking-300k benchmark and is releasing the model under an Apache 2.0 license on Hugging Face and GitHub.

Why this matters

This matters more than it may look at first glance. Many enterprises want to use generative AI in logging, search, indexing, and internal workflows, but get blocked by privacy and data risk. A compact model that can run locally and strip out PII before data moves deeper into the pipeline could become a practical building block for enterprise AI. It also strengthens OpenAI’s position in the security and infrastructure conversation, not just in the foundation-model race.

Source and date validation

The original source is OpenAI’s own research post, “Introducing OpenAI Privacy Filter,” published on April 22, 2026. OpenAI also linked the official model releases on Hugging Face and GitHub the same day. The date is official and within the 48-hour freshness window, so the story qualifies as current.

Source: https://openai.com/index/introducing-openai-privacy-filter/