Episode
OpenAI Open-Sources Privacy Filter, a Tiny Model That Scrubs PII Without an API Call
- Published
- Apr 23, 2026
- Duration seconds
- 621
- Processing state
not_requested- Canonical source
- https://share.transistor.fm/s/71ea78f5
Actions
POST https://stenobird.com/v1/public/podcasts/tech-companies-tech-brief-by-hackernoon-6365644/episodes/openai-open-sources-privacy-filter-a-tiny-model-that-scrubs-pii-without-an-api-call/transcription-requests
Idempotently request low-priority transcript generation for this episode.GET https://stenobird.com/podcast/tech-companies-tech-brief-by-hackernoon-6365644/openai-open-sources-privacy-filter-a-tiny-model-that-scrubs-pii-without-an-api-call.md
Read the agent-friendly Markdown representation of this episode resource.
Summary
This story was originally published on HackerNoon at: https://hackernoon.com/openai-open-sources-privacy-filter-a-tiny-model-that-scrubs-pii-without-an-api-call . OpenAI open-sourced Privacy Filter, a 50M-active-parameter model that detects and masks PII locally in one pass. Here's what's actually new, and what's hype. Check more stories related to tech-companies at: https://hackernoon.com/c/tech-companies . You can also check exclusive content about #openai , #open-source , #privacy-filter , #openai-privacy-filter , #tiny-model , #ai , #openai-open-sources , #hackernoon-top-story , and more. This story was written by: @abstraction . Learn more about this writer by checking @abstraction's about page, and for more stories, please visit hackernoon.com . OpenAI released Privacy Filter under Apache 2.0 — a 1.5B-parameter (50M active) bidirectional token-classification model that detects and masks PII in text locally, in a single forward pass. It runs on a laptop, supports 128K context, hits 96% F1 out of the box, and is fine-tunable with minimal data. Eight entity categories: names, addresses, emails, phones, URLs, dates, account numbers, and secrets. It's context-aware (not regex), ships with a CLI and eval tooling, and slots into the same open-weight ecosystem as gpt-oss. The catch: multilingual support is thin, adversarial formatting breaks it, and the benchmark validation used OpenAI's own models to grade OpenAI's model.