All articles

Tenant isolation is a clinical requirement.

A genomeSequencing of all of a person's DNA, coding and non-coding regions alike. identifies a person for life. A platform holding many services' cases has to keep each one's data apart at every layer, and be able to prove it.

Two laboratories in neighbouring regions use the same hosted interpretation platform. A scientist at one, searching for other carriers of a rare variant, is shown a match. The match is a patient of the other laboratory. The scientist sees a case identifier, a phenotype summary and a variant list before realising that the record is not one of theirs. Nothing was downloaded, nothing was printed, and the platform's logs, when they are examined, show the query but not what it returned.

The patient will never know. The two laboratories now have to decide whether that is the point.

What the field knows

Genomic data cannot be anonymised in the way other clinical data can. Homer and colleagues (2008) showed that an individual's presence in a pooled dataset could be inferred from aggregate allele frequencies. Gymrek and colleagues (2013) recovered surnames from Y-chromosome data in a public research set by cross-referencing genealogy databases. Erlich and Narayanan (2014) catalogued the routes by which genetic data are re-identified and the limits of each protection, and Wan and colleagues (2022) updated that review with the sociotechnical safeguards that have been tried since, concluding that technical controls, access governance and legal deterrents each cover part of the risk and none covers all of it.

The consequence for a clinical platform is that an exposure lasts. A leaked password can be rotated. A leaked genome identifies the patient and their relatives for as long as the data exist, and the harm, insurance, employment, family, is not reversible by a notification letter. The threshold for what counts as an acceptable isolation failure is therefore lower than for most hosted software.

The regulatory frameworks that apply, whichever jurisdiction a laboratory is in, share a requirement that data be accessible only to those with a reason, that access be recorded, and that the controller be able to demonstrate both. For a platform serving many laboratories, demonstrating it means being able to show, for any record, which organisation's users could ever have reached it, and that the boundary was enforced by the system rather than by convention.

Isolation also has a clinical purpose beyond privacy. Two laboratories may classify the same variant differently for good reasons, hold different consent terms for the same family, or be at different stages of validating a pipeline. A platform that lets one laboratory's decisions leak into another's analyses, through a shared cache, a shared default or a shared model, blurs responsibility for the result. Each service has to be able to say that a conclusion on its report was reached on its own data, under its own policies, by its own people.

Where practice breaks

Isolation is implemented once, at the front door, and assumed everywhere behind it. The application checks which organisation a user belongs to and then queries a shared database with that organisation as a filter. A single query without the filter, in a search feature, a report, an export or a background job, returns every organisation's data. The failure in the opening case is of exactly this kind, and it is the most common kind.

Shared features are built across the boundary by design. A matching service that finds similar cases, a variant frequency computed across the platform, a model trained on everyone's data: each is useful, and each is a path from one organisation's records to another's that has to be designed deliberately or it will be designed by accident.

Automated components inherit more access than any user. A pipeline worker or an assistant that reads cases on a user's behalf is often given a service credential with reach across the platform, because scoping it per request is harder. It then does whatever it is asked, including by text in an uploaded document that was never meant as an instruction.

The log shows that something happened and not what. An access log that records a query without the records it returned cannot answer the question that matters after an incident: whose data were seen. The laboratory cannot notify the right patients because it cannot name them.

Testing covers the happy path. Isolation is verified by checking that a user sees their own data, rarely by checking that every endpoint, every export and every background job refuses another organisation's. The negative test is the one that matters and the one that is written last, if at all.

How Tulparus handles it

Every tenant-owned row carries the tenant that owns it, and the database enforces the boundary itself through row-level security, so that a query without a tenant filter returns nothing. The tenant a request acts for is validated on every request against the person's actual memberships; it is never taken from a request body, a parameter or a claim the client supplies. A selection narrows what a membership already allows and never widens it.

Cross-tenant features are built as explicit, consented exchanges, with no shared tables. Similar-case matching between organisations happens only where both have opted in, on the minimum the match needs, and the match returns the fact of a match and a route to the other organisation's contact, never the other case. No patient data are used to train a model by default, and a tenant's data are not visible to any other tenant's analyses.

Automated components get no more reach than the person they act for. An agent or worker reaches services only through a gateway that applies the same tenant authorisation as a user's request, and text that arrives in uploaded or retrieved documents is treated as data rather than as instruction. There is no credential in the system that can read across tenants on a user's behalf.

Access is recorded at the level of the records returned, so that if a boundary ever fails, the question of whose data were exposed has an answer. The audit trail is itself tenant-scoped, and the organisation can export it for its own regulator. The scenario in the opening case cannot occur as described, and if a comparable failure did, the laboratory would know by the next morning which patients to tell.

What you can do tomorrow

These are questions to put to any platform that holds your cases, including one you run yourself.

  • Ask where the tenant boundary is enforced. If the answer is 'in the application', ask what happens to a query that forgets the filter.
  • Ask which features cross the boundary by design, and what a user of another organisation can see through each of them.
  • Ask what access an automated component or assistant has, and whether it is scoped per request or held as a standing credential.
  • Ask for the access log for one of your own cases, and check whether it shows which records were returned, or merely that a query ran.

Sources

  1. Homer N, Szelinger S, Redman M, et al. Resolving individuals contributing trace amounts of DNA to highly complex mixtures using high-density SNP genotyping microarrays. PLoS Genetics, 2008.
  2. Gymrek M, McGuire AL, Golan D, Halperin E, Erlich Y. Identifying personal genomes by surname inference. Science, 2013.
  3. Erlich Y, Narayanan A. Routes for breaching and protecting genetic privacy. Nature Reviews Genetics, 2014.
  4. Wan Z, Hazel JW, Clayton EW, Vorobeychik Y, Kantarcioglu M, Malin BA. Sociotechnical safeguards for genomic data privacy. Nature Reviews Genetics, 2022.

One platform, many services.See how the boundary is enforced.

A demo takes 20 minutes, screen shared, on synthetic cases. Bring the question from one of your own.