Compliance

Which HIPAA Safe Harbor Identifiers RedactProof Detects Automatically

Safe Harbor de-identification under 45 CFR 164.514(b)(2) means removing all 18 identifier categories - not the ones your software happens to catch. This page maps each identifier to RedactProof's automatic detection, and is equally clear about the categories that still need a manual pass. Verified July 2026.

By RedactProof Editorial Team · Jul 21, 2026 · 4 min read

Vendors rarely publish this mapping. We think that's backwards: if you're building a de-identification workflow on top of a tool, you need to know exactly where the tool stops and your reviewers start. What follows is the honest version, identifier by identifier, as of July 2026.

How to read this mapping

Detected automatically means RedactProof's pattern matching or on-device AI model flags the identifier as a suggestion, with a confidence score, for you to approve. Nothing is redacted without approval. Manual means detection does not analyze that content today - a reviewer draws the redaction box. For Safe Harbor work, treat every document as requiring human review regardless of what was auto-detected. Automated detection reduces reviewer workload; it does not replace the reviewer.

Detected automatically (identifiers 1-10)

The bulk of the list is covered by automatic detection:

1. Names. Detected by both pattern matching (labeled name fields, SURNAME, FIRSTNAME formats common in medical headers) and the AI model.
2. Geographic subdivisions smaller than a state. Street addresses, cities, and zip codes are detected. Remember Safe Harbor permits the first three zip digits only under population conditions - a decision for your reviewer, not the software.
3. Dates. Dates of birth are detected when labeled (DOB, Date of Birth, Born). Other Safe Harbor dates - admission, discharge, death - are not specifically classified, so sweep for them manually. Ages over 89 also need reviewer judgment.
4-6. Phone numbers, fax numbers, email addresses. All detected. Fax numbers share the phone number formats.
7. Social Security numbers. Detected with format validation (invalid area and group numbers are rejected, which cuts false positives).
8. Medical record numbers. Detected, including labeled accession numbers, specimen and sample IDs, pathology numbers, and hospital IDs - the identifier soup that fills real lab reports.
9. Health plan beneficiary numbers. Member, policy, subscriber, and group IDs are detected when labeled. Medicare Beneficiary Identifiers are validated against the CMS character rules.
10. Account numbers. Bank accounts, payment cards (checksum-validated), IBANs, and routing numbers are detected.

Detected with caveats (identifiers 11-15, 18)

11. Certificate and license numbers. Driver's license numbers are detected. Professional certificate numbers are caught when they appear with a label (License No, Certificate #); unlabeled ones need a manual sweep.
12. Vehicle identifiers. License plates and VINs are detected by the AI model.
13. Device identifiers and serial numbers. Detected when labeled (Serial No, S/N, Device ID, UDI, Equipment/Asset numbers). A bare serial with no label needs manual review.
14. URLs. Detected by the Precision engine. Pattern-based URL flagging is off by default because most URLs in documents aren't personal data - turn a careful eye to patient portal links.
15. IP addresses. Both IPv4 and IPv6 are detected.
18. Any other unique identifying number, characteristic, or code. Labeled reference, case, and claim IDs are detected, as are employee IDs. By definition this category is open-ended - no tool can claim complete coverage of it, and any vendor that does is overclaiming.

Manual redaction required (identifiers 16-17)

Two categories are not analyzed by automatic detection today. Plan your workflow around them.

16. Biometric identifiers. Fingerprints, voiceprints, and similar biometric content need manual redaction.
17. Full-face photographs and comparable images. Photographs of people are not detected as identifying content. Draw redaction boxes over them. The same applies to identifying images that aren't faces - a distinctive tattoo in a wound photo, for instance.

Related content that also lives outside detection: barcodes and QR codes are not decoded, handwritten signatures have no detection category, and images embedded on text-heavy pages need a per-page OCR re-scan before their contents are visible to detection. The full list is on the detection coverage and limitations page.

What a defensible Safe Harbor workflow looks like

Run automatic detection first. Review every suggestion - approve, dismiss, or adjust. Then do a manual pass focused on the gaps above: photographs, biometrics, unlabeled identifiers, barcodes, and Safe Harbor dates beyond birth dates. Export produces a pixel-burn PDF (the underlying text is destroyed, and no source metadata survives) plus a certificate recording what was redacted and whether each item was auto-detected or manually added. For the regulation itself, our Safe Harbor checklist guide walks through all 18 identifiers from the compliance side.

Disclaimer: This guide is for informational purposes only and does not constitute legal, medical, or professional advice. Consult a qualified professional for advice specific to your situation.

Frequently Asked Questions

Does RedactProof guarantee Safe Harbor de-identification?

No, and no software can. Safe Harbor requires removing all 18 identifier categories and having no actual knowledge that remaining information could identify an individual - that second part is inherently a human judgment. RedactProof detects most identifier categories automatically and records the review trail, but the de-identification determination sits with your organization. Two categories (biometrics and full-face photographs) always need manual redaction.

Are scanned documents and faxes covered?

Yes, with one caveat. Pages with little or no selectable text are OCR'd automatically before detection runs, so scanned records and faxes work. The caveat is mixed pages: an embedded image sitting on a page that already has typed text (a scanned wristband pasted into a report, say) needs a per-page re-scan before detection can read it. The editor flags these pages for you.

Does document content leave the browser during detection?

Documents are processed in your browser and are not uploaded to our servers. The standard detection engine runs entirely on-device. The optional Precision engine sends extracted text (not the file) to Cloudflare Workers AI for enhanced detection - processed in memory, not stored, not used for model training. For PHI workflows where text must not leave the network, use the standard engine.

Will the same document produce the same detections next year?

For a given model version, yes - the on-device engine is deterministic. But detection models improve over time and versions are not pinned per account, so a document re-processed after a model update may be detected differently. Certificates record which engines produced a document's redactions, so your audit trail survives model updates. Treat each document as reviewed at time of processing.