45 CFR Part 160 + 164

AI safety + compliance for healthcare

The eval + guardrail + red-team + audit platform purpose-built for AI workloads that touch PHI. Eliminate hallucinations on clinical decision support, redact PHI before it reaches third-party LLMs, generate the audit trail your auditors and payers expect.

Start free trial Request a BAA Book a demo

235

Scorers

LLM providers

334

Red-team plugins

2.57ms

Firewall p95

Clinician in a white coat with a stethoscope using a mobile device — Healthcare45 CFR Part 160 + 164

What ships today

Honest posture, not roadmap promises

Every checked item is in production today. In-progress items are flagged explicitly — no overclaiming, no vapor.

HIPAA-aligned technical safeguards (§164.312)

HIPAA-aligned administrative safeguards (§164.308)

AI-specific HIPAA controls (6 categories mapped)

Breach Notification Rule flow (§164.400-414)

Privacy Rule controls (§164.500-534)

BAA available on request (Business Associate)

Third-party HIPAA attestation by accredited firm

Built for buyer reality

Healthcare AI use cases we ship for

Chart summarization + clinical handoff

AI summarizes a patient chart for a covering clinician at shift change. PHI must stay inside the workload; summaries must not hallucinate a diagnosis or medication that isn't in the record.

EvalGuard features

PHI firewall: redacts SSN / MRN / DOB / patient name from logs by default
Hallucination scorer: faithfulness check against the source chart
Output guardrail: block mode rejects responses introducing un-cited claims
Audit log: every summary logged with clinician + patient resource_id for §164.312(b)

Patient-facing triage chatbot

Public-internet chatbot collects symptoms + routes patients to the right service. Cannot output medication dosing, cannot impersonate a clinician, cannot leak prior patients' PHI.

EvalGuard features

Prompt-injection defense: 300+ attack plugins catch jailbreak attempts trying to extract prior sessions
Output toxicity + medical-advice scorers flag responses crossing into clinical advice
Per-route rate limits prevent enumeration attacks against the symptom-collection endpoint
Consent gates at /api/v1/privacy/consent — refuses LLM call when subject withdraws

Prior-authorization automation

AI drafts prior-auth letters from claims + chart data, then a human reviews + signs. PHI flows through multiple LLM calls; cost + audit trail per claim must be perfect for payer disputes.

EvalGuard features

Gateway cost ledger: per-claim USD attribution + per-tenant daily budget cap
Tamper-evident audit log (integrity_hash + advisory-lock concurrency) for payer dispute defense
BYOK provider keys: routes through Vault-stored credentials, never logged
Multi-model routing with quality-cost strategy — cheapest model meeting accuracy threshold

Voice telemedicine guardrails

Live voice consult between clinician + patient is transcribed + summarized in real-time. Need voice-clone detection, barge-in latency under 250ms, deepfake guard on patient identity verification.

EvalGuard features

Voice-clone + deepfake guardrails on the realtime voice control plane
Per-frame guardrail scan via the gateway realtime WebSocket proxy
Barge-in latency budget enforceable with per-chunk stream scanning
Every verification attempt logged with signals for clinical review

Wire it in 60 seconds

Wrap your OpenAI client. Get PHI guardrails for free.

PHI firewall rules + clinical-faithfulness scorers + BAA-protected audit retention are configured once in the EvalGuard control plane. Your code only wraps the client.

typescript

import OpenAI from "openai";
import { wrapOpenAI, EvalGuardViolationError } from "@evalguard/openai";

const openai = wrapOpenAI(new OpenAI(), {
  apiKey: process.env.EVALGUARD_API_KEY!,
  projectId: "phi-chart-summary",
  metadata: { vertical: "healthcare", hipaa: true },
  blockOnViolation: true,                  // refuse PHI-leaking prompts
  evalOnResponse: { failOnScore: 0.7 },    // clinical-faithfulness gate
  onViolation: (r) => alertSecOps(r.violations),
});

try {
  await openai.chat.completions.create({
    model: "gpt-4o",
    messages: [{ role: "user", content: chartSummaryPrompt }],
  });
} catch (err) {
  if (err instanceof EvalGuardViolationError) {
    // Block landed in BAA-protected audit trail. Replay via audit ID.
  }
}

Firewall PHI patterns + scorer thresholds + retention live in the EvalGuard control plane — set once per project, no SDK calls needed.

Same integration for Anthropic, Gemini, and 91+ providers — swap wrapOpenAI for wrapAnthropic.

Stack