Privacy and security

Uploaded files are stored in a private object-storage bucket. They are not publicly reachable: there is no URL that serves them without authentication.

When someone in your workspace opens a document, Stride checks their membership first and then issues a link that expires after a few minutes. A forwarded link stops working shortly after it is created.

The extracted text is also stored in the database, split into passages, so that analysis can quote and cite specific parts of a document.

Documents are read to build your company profile and to ground strategy analysis in what your company actually knows, rather than in generic advice.

When Stride generates something, it retrieves the passages most relevant to the question and includes them in the request to the model. It does not send your entire document library on every request.

Stride does not use your documents to train any model. Nothing you upload is used to improve the product for other customers.

Generating analysis means sending the retrieved excerpts, your answers, and your company profile to an external model provider (Anthropic or OpenAI, depending on configuration).

What that provider does with those inputs is governed by their terms for the account whose API key is configured, not by Stride. Commercial API tiers generally do not train on inputs, but you should confirm this against your own agreement.

Stride can be run in mock mode, where no external model is called at all and no data leaves the machine it is running on. Check the AI provider section in Settings to see which mode you are in.

Only metadata is recorded for observability. Model name, token counts, timing. Prompt and completion text are not sent to any third-party monitoring service.

Everything is scoped to your workspace. Data is isolated at the database level, not only in application code: a query for one workspace cannot return another workspace's rows even if the application has a bug.

Within a workspace, all members can read the documents and the strategy work. Roles control what people can change: editors can upload and edit, admins and owners can approve diagnoses and strategy decisions and apply roadmap changes.

There is currently no per-document permission. Anyone in the workspace can read every document in it. If that matters for a particular file, do not upload it.

Stride records an audit event for security-relevant actions: uploads, downloads, deletions, approvals, overrides, exports, and membership changes. Each records who did it, to what, and when.

The audit trail is append-only. There is no interface, and no database permission, that lets anyone edit or delete it.

Only admins and owners can read the audit trail. It records who looked at what, which is itself sensitive.

Application logs deliberately exclude document content, problem statements, quoted excerpts, and credentials. Identifiers are logged; the confidential text behind them is not.

Deleting a document removes the stored file, its extracted passages, and every citation pointing at it. Analysis that previously cited it keeps its text, but the citation no longer resolves.

Superseded versions of diagnoses, roadmaps, and briefs are kept rather than deleted. This is deliberate: the record of what was decided, and on what basis, is the point of the product.

There is a development-only endpoint that deletes an entire workspace and everything in it. It is disabled unless explicitly enabled by an environment variable, and it is not available in production.

A document can contain text designed to manipulate a language model: instructions telling it to ignore its rules or to report something false.

Stride treats all document content as data, never as instruction. Retrieved passages are wrapped in an explicit boundary, tags that could be used to break out of that boundary are stripped, and the model is instructed that content inside it is evidence to analyze rather than commands to follow.

Citations are verified against the database before being stored: a source the model invented resolves to nothing and is dropped, and the claim it supported is downgraded from "from your documents" to "inferred".

No invitation emails. Team members must sign in once before they can be added to a workspace.

No OCR. A scanned PDF with no text layer is marked unsupported rather than being silently indexed as empty.

No per-document permissions, and no way to exclude a document from analysis other than deleting it or scoping an initiative to specific documents.

No customer-managed encryption keys, and no data residency controls.

No penetration test and no third-party security audit has been performed.