OpenAI open-sources Privacy Filter
- OpenAI released Privacy Filter as open-source to mask sensitive information before text hits chatbots. - The model strips personally identifiable information so secrets never reach the underlying AI model. - Treating privacy-preserving preprocessing as a product feature helps vendors sell AI into regulated sectors. (decrypt.co)
OpenAI has released Privacy Filter, an open-weight model that strips sensitive details from text before that text reaches a chatbot. (openai.com) The company published the model on April 22, 2026, and made the code available in a public GitHub repository under the Apache 2.0 license. OpenAI said developers can run it in their own environments and fine-tune it for training, logging, indexing, and review pipelines. (openai.com) (github.com) Privacy Filter is built to read a full passage at once and label the parts that should be hidden, instead of generating text word by word. OpenAI says that single-pass design lets teams process long documents quickly on local machines. (github.com) (openai.com) That matters for companies that do not want raw client records, internal emails, or credentials sent to an outside server just to be cleaned up. OpenAI said the model can run locally, so unfiltered text can stay on the device while names, account numbers, and other identifiers are masked first. (openai.com) Older privacy tools often work like rigid templates, catching obvious patterns such as phone numbers or email addresses but missing context. OpenAI said Privacy Filter is meant to handle unstructured text and decide when a detail refers to a private person and when similar information is already public. (openai.com) The repository describes eight output categories, including names, addresses, emails, phone numbers, Uniform Resource Locators, dates, account numbers, and secrets such as passwords or application programming interface keys. In OpenAI’s example, the model swaps those details for placeholders like `[PRIVATE_EMAIL]` and `[ACCOUNT_NUMBER]`. (github.com) (decrypt.co) OpenAI says the model has 1.5 billion total parameters, with 50 million active parameters, and supports a 128,000-token context window. The company says it is small enough to run in a web browser or on a laptop, which lowers the hardware bar for teams that want on-premises filtering. (github.com) OpenAI said the release version reached state-of-the-art performance on the PII-Masking-300k benchmark after correcting annotation issues it found in that evaluation set. Decrypt reported the out-of-the-box score at 96% F1, with a corrected version of the same test reaching 97.43%. (openai.com) (decrypt.co) The model card also warns against treating the system as a complete privacy solution. OpenAI lists over-reliance, failure modes, and high-risk deployment concerns, and recommends human review and broader controls around sensitive data. (cdn.openai.com) OpenAI said it already uses a fine-tuned version of Privacy Filter in its own privacy-preserving workflows. By open-sourcing the same approach, the company is pushing a tool that lets enterprises scrub text before an underlying language model ever sees the original secrets. (openai.com)
