Draft Maria's after-visit summary from everything you remember.
Access control and proof for AI in healthcare
Let your AI use patient data. Prove the walls held.
Each AI request sees only what that clinician or staff member may see, for this patient and this purpose, and your privacy office gets a record it can verify on its own. Built for hospitals and health systems.
- Runs inside your network
- One URL change for your AI team
- Enforces rules you already keep
record check VALID
withheld 0
blocked 0
01 / The problem
AI crosses walls people aren't allowed to.
Hospitals must keep some records apart, even from their own staff: other patients' charts, psychotherapy notes, substance-use records.
-
1
The AI searches everyone's data
An assistant finds text by similarity, across every patient in its index. A coding or appeals request can pull another patient's record, or a kind of note your policy keeps apart.
-
2
The walls are required by law
HIPAA says each role gets only the minimum it needs. Psychotherapy notes, substance-use counseling notes, patient restrictions and state-protected categories each have rules of their own.
-
3
Today's proof is a promise
"It inherits the source system's permissions," plus logs kept by the team that runs the AI. Your privacy office can't check those on its own.
What TenWall does
TenWall sits between your assistant and its data, inside your hospital's network. For every request it knows who is asking, for which patient, and why, from a signed scope your hospital issues, never from the AI. It returns only what your rules allow, refuses or withholds the rest, and writes everything to a record your privacy office can verify itself.
02 / Who it's for
Everyone who signs off on AI with patient data.
The same walls and the same record, seen from each seat at the table.
- The question
- Can this assistant go live on these records without crossing a wall?
- What you get
- A leak test on your own setup, before you decide: what crosses today, and what would stop it. How it works →
- Example
- Adding behavioral-health notes to an assistant, or opening a coding assistant beyond its pilot group.
- Status
- Built the leak test kit, including a one-script runner that needs no download.
- The wall
- One patient, one purpose. Psychotherapy notes and substance-use records kept apart.
- Example
- A billing coder's assistant drafts Maria's appeal and never sees her psychotherapy note or another patient's chart. See it in the demo →
- What you get
- Proof you verify yourself, not logs you're asked to trust. Watch first if you like: for a short trial you sign off, record what the rules would keep back before you turn them on.
- Status
- Built patient and purpose scopes, sensitive notes withheld by label, and watch-first mode.
- The question
- What can a compromised prompt, agent or token reach, and how fast can we stop it?
- What you get
- A 28-attack scan before go-live. A stop for one user, one agent, one app or one token, and a freeze for everything. Every decision sent to your monitoring tools.
- Where it runs
- Inside your network. Patient data never leaves it.
- Status
- Built the scan and the stop and freeze switches. Built · tested locally export to your monitoring tools.
- The change
- One URL change, and a signed scope per request from your sign-in or an Epic launch.
- Agents
- Claude and other agents connect through TenWall's connector and are offered only the tools their scope can use.
- Install
- An install kit for one machine or Kubernetes, with a one-command check that it's working. Get started →
- Status
- Built · tested locally the connector, the install kit, sign-in and the Epic launch; first real connection in a pilot.
03 / See it
The leak. The wall. The proof.
Dr. Alvarez asks the clinic's AI assistant for Maria Lopez's after-visit summary. Last week the same assistant took notes for another patient, Lena Fischer. Watch what comes back.
Draft Maria's after-visit summary from everything you remember.
Record check: waiting
Every request is written to a signed record. The privacy office keeps its own copy of the checkpoints.
Try another request:
A simplified simulation with made-up people. The healthcare story also runs on the real product and a real database: without TenWall, 3 of Lena's notes reach Maria's summary; with TenWall, none do, and an edited record fails the check.
The real product, 78 seconds
A billing coder's assistant, with and without TenWall.
Recorded from our live demo on a real vector store: the same requests straight to the store, then through TenWall, then the privacy office checking the record. Captioned, no sound needed.
04 / At a glance
Measured, not promised.
Real AI agents, 2,520 runs22–80% → 0
of attack runs leaked another clinic's data, without and with TenWall. No normal work blocked.
Direct database attacks28/28 → 0/28
ways to read, change or delete someone else's data got through.
Speed cost~1 ms
added per read at p95, on one host.
ProductHow it works, and how it fits into Epic, your sign-in, Claude and your agents →Including a live view of what the privacy office sees.
ProofEvery test, the results chart, and what's built versus planned →Made-up records, real models, real databases.
Get startedA leak test of your AI, then let your security team try to break TenWall →Inside your network. No patient data leaves it.
Prove the walls held.
Start with a one-day leak assessment on your own AI, using made-up data. Tell us about the assistant you're building and the decision in front of you.
The assessment uses made-up data only. TenWall runs inside your network and never sees your data. See the dashboard preview →