OpenAI has released Privacy Filter, a small open-weight model designed to find and remove personal information from text before it is sent to a cloud AI service or stored. The company published the model this week under an Apache 2.0 license on Hugging Face and GitHub, making it free for developers to use, modify and deploy.
The release addresses a persistent problem for organizations adopting AI tools: employees and customers routinely paste names, contact details, account identifiers and other sensitive data into prompts, where it can end up in logs, training pipelines or third-party systems.
How the model works
Privacy Filter has about 1.5 billion parameters, small enough to run on a laptop or inside a web browser without specialized hardware, according to OpenAI. It scans text and labels spans that fall into eight categories of personally identifiable information, including personal names, email addresses, phone numbers, physical addresses and account or identity numbers. Applications can then mask or delete those spans before the text goes anywhere else.
Because the model runs locally, OpenAI says, the raw data never needs to leave the user’s device for the redaction step itself. That design is aimed at regulated sectors such as healthcare, finance and law, where organizations want the benefits of large cloud models but face strict rules about what data may be transmitted.
Benchmark results and limitations
OpenAI reports that Privacy Filter achieves an F1 score of about 96 percent on PII-Masking-300k, a public benchmark for detecting personal data in text. F1 is a combined measure of how many sensitive items a system catches and how often it flags text incorrectly, so a high score indicates both good coverage and few false alarms.
The company acknowledges limits. Performance on languages other than English is weaker, and OpenAI describes multilingual support as an area for further work. The model is also a classifier for text only; it does not handle images, audio or documents with complex formatting unless they are first converted to plain text.
Why an open release
Privacy Filter joins a small but growing set of open-weight releases from OpenAI, a company that keeps its flagship models proprietary. A privacy tool is a natural candidate for open distribution: organizations are more likely to trust a redaction component they can inspect and run entirely under their own control, and wide adoption of such tooling lowers the risk that sensitive data reaches OpenAI’s own services.
Developers can download the weights from Hugging Face or the GitHub repository, where OpenAI has published inference code and documentation for integrating the model into existing data pipelines.
