Skip to main content
Healthcare AI Security

Your Patients' Data Stays on
Your Machines.

Doctors, nurses, and billing teams are using ChatGPT, Claude, and Copilot every day — pasting patient names, medical record numbers, and insurance IDs into prompts. Shield sits on your hospital's machines and stops protected health information from ever reaching an external AI model. No cloud. No Business Associate agreement needed. Just a local proxy that redacts PHI before it leaves your network.

See Shield for HealthcareSee PHI Redaction in Action
$7.42M
Average healthcare breach cost

Healthcare has held the highest breach cost of any industry for 14 consecutive years. The average healthcare breach costs significantly more than any other sector, driven by regulatory penalties and long detection times.

Source: IBM Cost of a Data Breach Report 2025
18 Identifiers
PHI data elements defined by HIPAA

Names, dates, SSNs, medical record numbers, IP addresses, biometric data, and more. Any one of these in a prompt creates a compliance exposure if it reaches an external AI model.

Source: HHS.gov HIPAA Privacy Rule
Zero
Business Associate agreements needed

Because Shield runs locally on your infrastructure, it never stores, transmits, or processes PHI on Purfect Labs servers. No BAA required — your compliance scope doesn't expand.

Shield operates entirely within your network boundary

HIPAA penalties escalate fast

HIPAA civil money penalties are tiered by culpability. Even a single PHI exposure caused by an employee pasting patient data into an AI tool can trigger an OCR investigation. The Department of Health and Human Services has issued resolution agreements totaling millions of dollars against healthcare organizations. Shield eliminates this vector entirely — PHI can't leak through AI prompts because it never leaves your network in the first place.

Interactive Demo

What PHI Looks Like in AI Prompts

Three real-world scenarios where healthcare staff send PHI to AI models. Click each tab to see the raw prompt with identifiable patient data — and how Shield redacts it before it leaves your network.

Doctor Drafting Clinical Notes with AI

A physician uses an AI assistant to draft a referral letter, pasting the patient's recent lab results and medical history into the prompt.

Before Shield
Please draft a referral letter for Jane Morrison (DOB: 03/14/1962, MRN: 48291-553), who presented with elevated HbA1c of 8.7% and fasting glucose of 162 mg/dL. Her current medications include metformin 1000mg BID and lisinopril 10mg daily. Referring to Dr. Sarah Chen, Endocrinology, at City General Hospital, fax 415-555-0182.
After Shield
Please draft a referral letter for [PATIENT_NAME] (DOB: [DATE], MRN: [MEDICAL_RECORD]), who presented with elevated HbA1c of [LAB_VALUE] and fasting glucose of [LAB_VALUE]. Her current medications include [MEDICATION] and [MEDICATION]. Referring to [DOCTOR_NAME], Endocrinology, at [FACILITY_NAME], fax [PHONE].
Detected PHI (6 matches)
PHI — NameJane Morrison
PHI — Date03/14/1962
PHI — ID48291-553
PHI — ProviderDr. Sarah Chen
PHI — LocationCity General Hospital
PHI — Contact415-555-0182

Compliance

How Shield Maps to HIPAA Security Rule

The HIPAA Security Rule establishes national standards for protecting electronic protected health information. Shield's local proxy architecture satisfies multiple Security Rule requirements without adding cloud dependencies to your compliance scope.

HIPAA Security RuleRefRequirementHow Shield Satisfies It
Access Control§164.312(a)(1)Implement technical policies and procedures for electronic information systems that maintain PHI to allow access only to authorized persons or software programs.Shield runs locally with OS-level access controls. Only authorized users on the host machine can start or configure the proxy. No remote access to redaction engine.
Audit Controls§164.312(b)Implement hardware, software, and/or procedural mechanisms that record and examine activity in information systems that contain or use PHI.Shield's tamper-evident audit trail records every prompt sanitization event with cryptographic hashing. Every redaction — what was caught, when, by which filter pack — is logged with a verifiable chain of custody.
Integrity§164.312(c)(1)Implement policies and procedures to protect PHI from improper alteration or destruction.Shield's redaction is non-destructive — the original prompt is never modified in-place. Redacted values are replaced with placeholders, and the mapping stays on your machine. Audit logs track every transformation.
Person or Entity Authentication§164.312(d)Implement procedures to verify that a person or entity seeking access to PHI is the one claimed.Shield integrates with your existing SSO and identity provider. Only authenticated users can configure filter packs, view audit logs, or modify redaction policies.
Transmission Security§164.312(e)(1)Implement technical security measures to guard against unauthorized access to PHI being transmitted over an electronic communications network.Shield redacts PHI before transmission. The data that leaves your network contains no identifiable patient information — only placeholders. Even if the transmission were intercepted, no PHI is exposed.

Architecture

PHI Never Leaves Your Network

Shield runs as a local proxy on your hospital's machines. When a clinician, biller, or administrator sends a prompt to an AI model, Shield intercepts it — redacts all PHI — and only then forwards the clean prompt to the external LLM. The redaction mapping stays on your machine.

🏥 HospitalEHR / PACS / LISprompt + PHI🛡️ Shieldlocal proxy✗ PHI caught & redactedclean promptLLM APIChatGPT / Claude🔒 PHI never leaves your hospital networkYour Network

Local Installation

Install Shield on any Mac, Windows, or Linux machine. No cloud infrastructure, no vendor data access, no data leaves your network.

One Environment Variable

Set SHIELD_PROXY_URL and every AI call from that machine flows through Shield automatically. Zero code changes to your existing tools.

Full Audit Trail

Every redaction event is logged with a cryptographic hash. Prove to auditors exactly what PHI was caught and when — with tamper-evident integrity.

Configurable Policies

Choose which PHI categories to redact. Add custom patterns for your facility's internal coding systems. Run in audit-only mode to validate coverage first.

Ready to Protect Patient Data in AI Workflows?

Shield runs on your hospital's existing infrastructure. No cloud. No BAA. No PHI leaves your network. Deploy in under an hour and give your clinical teams the AI tools they need — safely.

Shield for HealthcareTalk to Our Team

Frequently Asked Questions

No — and that's the point. Shield runs locally on your machines, not as a cloud service. It never sees, stores, or transmits PHI. Because Shield operates entirely within your own infrastructure and never sends data to Purfect Labs servers, it falls outside the HIPAA Business Associate definition. Your PHI stays on your network the entire time.
Yes. Shield's filter packs catch structured identifiers (SSNs, MRNs, NPI numbers, ICD-10 codes, CPT codes, dates of service) and unstructured PHI (doctor's notes, patient messages, referral letters). The secret detection engine uses entropy analysis to catch custom hospital IDs, internal patient keys, and proprietary coding systems that regex patterns miss.
Shield supports configurable redaction policies. For research workflows, you can configure Shield to apply the HIPAA Safe Harbor method — stripping all 18 PHI identifiers before prompts reach external models. This lets your research team use AI tools while maintaining compliance with the HIPAA Privacy Rule's de-identification standard.
Shield operates at the API layer, not the EHR layer. It doesn't integrate directly with any EHR — it sits between your AI client (whether that's a custom app, a browser extension, or a CLI tool) and the LLM provider. As long as the AI request flows through Shield's proxy, PHI is caught regardless of which EHR generated the data.
Shield combines multiple detection layers: regex patterns for known formats (SSN, MRN, NPI), entropy analysis for custom identifiers and encoded data, and configurable custom patterns for your organization's specific ID schemes. You can also run Shield in audit-only mode to log what would have been caught without blocking, letting you validate coverage before enforcing redaction. For maximum safety, Shield supports a deny-by-default posture — any prompt that can't be fully verified is blocked.