OpenAI Privacy Parser

OpenAI shipped Privacy Filter — a model that hides PII in text. Defenders use it so data doesn't leak.

But every defense has a second side. Attackers need PII too. From logs, dumps, abandoned S3 buckets, stolen inboxes. Same task: find everything personal in a pile of text.

Privacy Filter masks. Privacy Parser extracts.

Same model, same label taxonomy, same weights. Instead of <REDACTED> you get structured spans — what, where, which type.

Defense: audit your data before it leaks. Offense: parse someone else's after it did.

The tool doesn't know whose side it's on. It's just a good parser.

---

Install

uv venv
uv pip install -e ./privacy-filter
uv pip install -e ./pii_parser

First run downloads the opf 1.5B checkpoint (~3 GB) to ~/.opf/privacy_filter/.

Use

from pii_parser.hybrid import HybridPIIParser

parser = HybridPIIParser(device="cpu")
result = parser.parse(
    "Hi Quindle Testwick (quindle.testwick@openai.com / +1-415-555-0102), "
    "account 40702810500001234567, 14 Beautiful Ct, Anytown USA, "
    "password Priv4cy-Filt3r-2026."
)
for s in result.spans:
    print(f"{s.label:18}  {s.text}")

private_person      Quindle Testwick
private_email       quindle.testwick@openai.com
private_phone       +1-415-555-0102
account_number      40702810500001234567
private_address     14 Beautiful Ct, Anytown USA
secret              Priv4cy-Filt3r-2026

CLI

python -m pii_parser.cli_model "Alice paid 40702810500001234567 on 2026-05-17."

Three backends

Backend	Weights	Speed	F1
PIIParser	none	µs	1.000
ModelPIIParser	1.5B	500 ms CPU	0.733
HybridPIIParser	1.5B	600 ms CPU	0.929

Hybrid = model + span-merge + regex backstop. Ship this one.

Labels

opf v2 taxonomy — 8 categories:

private_person · private_email · private_phone · private_address · private_url · private_date · account_number · secret

Architecture

text
  ↓
opf 1.5B → BIOES logits → Viterbi (tuned) → char spans
  ↓
span-merge (glues Quindle + Testwick)
  ↓
regex backstop (URL, secret, account — where the model misses)
  ↓
spans[]

Benchmarks

python pii_parser/tests/test_hybrid.py

Fixture F1:  0.929
Scenarios:   8/8 passed
Latency:     ~600 ms CPU

License

Apache-2.0.