Krux
April 29, 2026
OpenAI's Privacy Filter Catches PII Before It Hits Servers
Published: April 29, 2026 at 12:19 AM
Updated: April 29, 2026 at 12:19 AM
100-word summary
OpenAI released Privacy Filter on April 22, a 1.5-billion-parameter model that spots eight types of personal data in documents with 96% accuracy. The twist: it runs on your laptop, not in the cloud. That means you can scrub phone numbers, emails, and addresses from customer tickets or support logs without sending anything to a server first. The model handles up to 128,000 tokens at once (roughly a short novel), so it won't choke on long transcripts. Fine-tuning on a handful of domain-specific examples lifted accuracy from 54% to 96% in OpenAI's tests. One catch: it's a detector, not a compliance stamp. You still need human review for high-stakes redactions.
What happened
OpenAI released Privacy Filter on April 22, a 1.5-billion-parameter model that spots eight types of personal data in documents with 96% accuracy. The twist: it runs on your laptop, not in the cloud. That means you can scrub phone numbers, emails, and addresses from customer tickets or support logs without sending anything to a server first. The model handles up to 128,000 tokens at once (roughly a short novel), so it won't choke on long transcripts. Fine-tuning on a handful of domain-specific examples lifted accuracy from 54% to 96% in OpenAI's tests. One catch: it's a detector, not a compliance stamp.
Why it matters
You still need human review for high-stakes redactions.