HIPAA & regulated workloads
Summary
Infer is designed for HIPAA-regulated workloads, one engagement in Infer's broader compliance program. The foundation is our zero-data-retention architecture: prompts and completions, including any protected health information (PHI) they contain, exist only in volatile memory for the life of a request and its session cache, are never written to durable storage, never logged, and never accessible to Infer personnel. Under HIPAA, the hardest questions are usually about what a vendor stores and who can see it; Infer's answer to both is nothing, and no one. That is by construction, not by policy alone.
Infer is pre-launch and does not yet accept PHI. Do not send PHI through Infer until a Business Associate Agreement (BAA) covering Infer is in place and we have confirmed in writing that our HIPAA program is active. This page describes the architecture and the program being built on top of it; it is not a certification claim; no government HIPAA certification exists for anyone to hold.
Where Infer sits under HIPAA
HIPAA obligations attach to covered entities (providers, plans, clearinghouses) and to the business associates that handle PHI on their behalf. Infer does not serve patients or covered entities directly; inference requests reach us through our routing partner. When a workload involves PHI, Infer acts in the role of a subcontractor business associate, and a BAA chain must connect the covered entity to every party that touches the PHI, including Infer and any infrastructure subprocessor on which the workload runs.
Customers remain responsible for their own HIPAA obligations, including whether an LLM workload involving PHI is appropriate for their use case and what their own policies permit.
How the architecture maps to HIPAA's safeguards
The HIPAA Security Rule asks for administrative, physical, and technical safeguards for electronic PHI. Several of its hardest technical requirements are addressed structurally by how Infer serves inference:
| HIPAA concern | Infer mechanism |
|---|---|
| Storage & disposal | Content is never written to durable storage, so there is nothing to retain, back up, sanitize, or dispose of. Volatile state is evicted automatically and does not survive a replica restart. |
| Transmission security | Traffic from our routing partner to the gateway is TLS-encrypted; traffic from the gateway to compute nodes runs over an encrypted overlay network. |
| Access control | A single authorized API caller, credentialed and optionally IP-allowlisted. No internal tool, console, or export path surfaces customer content to personnel. |
| Audit controls | Structured operational logs (timestamps, token counts, status codes, latency) record system activity without ever containing content. |
| Minimum necessary | Content is processed solely to generate the response. It is never inspected, mined, used for training, or reused for any secondary purpose. |
What zero retention does not do by itself
We are deliberate about the boundary of this claim. Architecture minimizes the risk surface; it does not by itself complete a HIPAA program. Before Infer accepts PHI, the following must be, and is being, put in place:
- Business Associate Agreements. An executed BAA with the party routing PHI to us, and BAAs covering any infrastructure subprocessor whose hardware processes PHI in memory. PHI-eligible capacity will run only on infrastructure covered by a BAA.
- A documented risk analysis and the written policies and procedures the Security Rule requires: access management, workforce security and training, incident response, contingency planning, and sanctions.
- Breach notification procedures. Zero retention shrinks the window in which a breach of content is possible, but the obligation to detect, assess, and notify exists regardless, and is documented and exercised.
BAA & documentation availability
A standard Business Associate Agreement is currently in legal preparation and will be available for qualifying workloads when the program activates. Infrastructure subprocessors are currently undergoing compliance evaluation for BAA coverage; see the security overview for how that evaluation works. Supporting security documentation is in preparation and will be made available to qualifying customers on request.
Before sending PHI
A workload may involve PHI on Infer only when all of the following are true:
- A BAA chain covering Infer (and our infrastructure subprocessors) is fully executed.
- Infer has confirmed in writing that its HIPAA program is active and PHI is accepted for your workload.
- Your own risk analysis covers the use of hosted LLM inference for the data involved.
To start that conversation, or to request our security documentation as it becomes available:
contact@infer.industries