meilynx

Framework · HIPAA

Security Rule evidence for the AI that touches PHI.

The Security Rule has bound covered entities since 2005 and business associates directly since 2013; the proposed modernization published in January 2025 is not final. Meilynx evidences the AI-traffic slice: which protected health information reached which model and what was redacted before the call, whether the approved AI configuration changed, and a tamper-evident, write-once record of every request. Detection and redaction support, never de-identification, never a certification.

What the rule expects

Safeguards, minimum necessary, associates, breach.

The technical safeguards of §164.312 apply to every information system that uses ePHI, the Privacy Rule's minimum-necessary standard applies to every prompt, and an LLM provider that receives PHI is a business associate.

  • Minimum necessary for every use and disclosure of PHI, with a standard protocol for routine disclosures (§164.502(b), §164.514(d)). An AI prompt is a disclosure to the provider behind it.
  • Technical safeguards: audit controls, access control, transmission security, integrity, and encryption for the systems that use ePHI (§164.312), retained six years (§164.316).
  • Administrative safeguards: an accurate and thorough risk analysis covering the AI systems that receive PHI, workforce training, and business associate contracts with every LLM provider (§164.308, §164.314).
  • Breach notification without unreasonable delay and within 60 days of discovery, with procedures that cover AI-workflow incidents (§§164.400–414).
How Meilynx maps

Runtime expectations, to runtime evidence.

A specific Meilynx control for each expectation the proxy can substantiate, and the artifact it produces. Everything else is attested in the package, and the package says which is which.

HIPAA Security Rule → Meilynx controls (AI-traffic slice)

Limit PHI in AI requests and responses to the minimum necessary

HIPAA · §164.502(b) · §164.514(d)

Maps to · The PHI detection rule scans every request and response for the Safe Harbor identifier classes it was benchmarked on and redacts each span in flight, before the request leaves your perimeter. Findings carry placeholders only; the control plane never holds PHI.

Examination artifact · PHI detection and redaction findings by workflow and provider

Record and examine activity in the systems that use ePHI

HIPAA · §164.312(b) · §164.316(b)(2)(i)

Maps to · Each AI request, response, and governance decision is sealed into a hash-chained record with write-once retention under your own key. Completeness across all systems that use ePHI is attested; tamper-evidence of the recorded events is verified.

Examination artifact · Hash-chained audit trail with six-year retention lock

Encrypt ePHI in transit and at rest

HIPAA · §164.312(e)(1) · §164.312(a)(2)(iv)

Maps to · TLS on every hop of the AI request path and customer-managed keys on the audit record. Encryption of your own data stores outside the proxy path is attested.

Examination artifact · Encryption posture with the cited mechanisms

Protect the AI configuration from unapproved alteration

HIPAA · §164.312(c)(1)

Maps to · Prompt-drift and tool-grant-drift detection flag changes to PHI-handling workflows against approved baselines; the governance bundle the data plane acknowledges is compared with the approved one.

Examination artifact · Drift findings and configuration-integrity verdict

Know which business associates received PHI, and attest the rest

HIPAA · §164.308(b)(1) · §164.314(a) · §164.308(a)(1)(ii)(A)

Maps to · The providers observed at the proxy are the LLM business associates the BAA control must cover. The BAAs themselves, the risk analysis, workforce training, and breach procedures are your program, attested in the package with structured evidence.

Examination artifact · Observed-provider inventory and administrative-safeguard attestations

The evidence

What you hand an OCR investigator.

The audit trail renders into a Security Rule evidence package for the AI-traffic slice. Each control is classified as proxy-verified runtime evidence or attested in your program, and the package says which.

In the package

  • PHI detection evidence: rule run-state, guard-model tier, redaction and fail-closed counts, placeholders only.
  • Audit controls, retention, encryption, access, and integrity evidence blocks with the cited mechanisms.
  • The observed LLM provider inventory the business associate agreements must cover.
  • Administrative safeguards: risk analysis, BAAs, workforce training, breach procedures, attested with structured fields.
  • Obligation timeline: Security, Privacy minimum necessary, and Breach Notification Rules in effect; the January 2025 modernization proposal as a watch item.
FAQ

HIPAA Security Rule and the proxy.

Is PHI detection de-identification?

No. The guard model detects the Safe Harbor identifier classes it was benchmarked on and redacts each span before the request leaves your perimeter. The package makes no Safe Harbor or expert-determination claim under §164.514(b). On the ratified benchmark it met the signed-off bar on every scored identifier class; license and vehicle numbers carry no benchmark support and are disclosed as unscored rather than claimed.

Does the preset make us HIPAA compliant?

No runtime product could. The preset evidences the AI-traffic slice of the Security Rule: six proxy-verified controls plus attestations for the risk analysis, business associate agreements, workforce training, and breach procedures. Compliance is a conclusion your privacy officer and OCR reach; the package is the evidence they read.

What happens when the PHI guard model is unavailable?

The authored disposition applies, block by default, and every occurrence is recorded as a critical finding. The package counts them. Nothing degrades silently.

What about the proposed Security Rule update?

The notice of proposed rulemaking was published on 6 January 2025 and the federal agenda targets July 2027 for final action. Nothing in the package relies on it; the obligation timeline lists it as a watch item.

Examination package

See exactly what an examiner receives

Download a sample examination package: model inventory, control coverage, a governance policy snapshot, and a SHA-256 integrity hash.

Regulatory updates

When a regulator changes what an AI examination asks for, hear about it first.

Short notes on SR 26-2, NYDFS 500, FINRA, the NAIC bulletin, the EU AI Act, and the employment-AI statutes, plus what we ship. A few emails a month.