Security
Security for private research knowledge
Researcher Center is designed for labs that need useful AI without losing control of unpublished work. The posture is deliberately conservative: evidence stays attributable, access stays bounded, and limitations stay visible.
Who reviews this page
Principal investigators, research IT, security reviewers, lab managers, and institutional buyers evaluating Early Access.
Privacy posture
Research documents, chunks, embeddings, prompts, and answers are treated as sensitive research data.
- The product is built around lab-owned corpora, not public-web answering.
- Managed cloud is optional; BYO-key and local/self-hosted paths are available for stricter review.
- Known limitations are kept visible so labs can decide what is appropriate to upload first.
Access boundaries
Organizations and workspaces define who can search, ask, and see sources.
- Tenant isolation is a core boundary: one lab's documents must not be retrieved for another tenant.
- Workspace-level access control scopes document retrieval and answer generation.
- Private app routes and backend APIs rely on server-side authorization, not browser-only checks.
Logging and auditability
Operational visibility should not become accidental data exposure.
- Sensitive document text is not written to default logs.
- Security-sensitive actions are designed to leave reviewable audit events.
- Provider failures and insufficient-evidence states should be visible without exposing research content.
Controls to review
- No cross-tenant document retrieval.
- No sensitive document text in default logs.
- Workspace permissions gate source access and answer generation.
- Audit events are recorded for sensitive operations.
- Security posture and Early Access limitations remain available for review.
Before uploading sensitive documents, review the known limitations and confirm which AI operating model your lab can approve.
Start with a corpus your lab already trusts
A focused pilot is the cleanest way to evaluate citations, refusal behavior, access boundaries, and deployment fit.