Section 1
What we will and will not do.
The plainest statement of scope a regulated-industry product can publish. Five lines on each side, no fine print.
We will
- ✓Isolate every org's data with Postgres Row-Level Security at the database layer
- ✓Encrypt your data in transit (TLS) and at rest on our infrastructure providers
- ✓Run drafted claims through the citation, off-label, and PHI guardrails that are live
- ✓Give you an Ed25519-signed, independently verifiable audit export per Vault
- ✓Give 30 days notice before any sub-processor change takes effect
We do not
- ✕Train any model on your data
- ✕Share your prompts, documents, or outputs across tenants
- ✕Send your text to any provider other than Anthropic and Voyage
- ✕Generate medical claims about your products or competitors
- ✕Predict trial outcomes or write MLR-final copy
Section 2
Certifications & attestations.
Honest status. Nothing here is certified yet. In progress means active work toward it. Planned means scoped but not started. We will not list a cert as achieved until the report is in hand and we can show it under NDA.
Operational security audit
Working toward it. Not yet certified.
Information security management
Planned. Not yet certified.
AI management system
Planned. Not yet certified.
Data Processing Agreement
Available on request: privacy@krux.bio
Business Associate Agreement for PHI
In progress. Ask under NDA: privacy@krux.bio
Electronic records & signatures
Krux is a proposal-drafting tool, not your GxP system of record. See note below.
On 21 CFR Part 11: Krux helps prepare proposal content. The regulated record of submission lives in your own validated systems, not in Krux. We do not position Krux as your GxP system of record. If your use case needs Part 11 controls on the records Krux holds, talk to us at compliance@krux.bio before you contract so we can scope it honestly.
Section 3
Architecture.
The security properties of the data path: tenant isolation, per-Vault access, encryption, and audit integrity.
Hosting
Application compute on Vercel. Postgres database, authentication, and object storage on Supabase. Both run on AWS underneath.
Tenant isolation
Postgres Row-Level Security. Every org-scoped table filters on the caller's org, enforced at the database layer, not in application code. New accounts get a private org at signup.
Per-Vault access control
Inside an org, a Bid Vault has its own role membership (owner, editor, reviewer, observer). Belonging to the org does not grant access to a Vault; a user must be added to it.
LLM data flow
Prompts and retrieved context go to Anthropic at inference time. Document chunks go to Voyage at embed and rerank time. Those are the only two recipients of customer text outside the platform.
Audit log integrity
Each Bid Vault keeps a SHA-256 hash-chained activity log. Exports are Ed25519-signed and can be verified in the browser against the public key registry, with no Krux server in the loop.
Encryption
TLS in transit. Encrypted at rest by the Supabase and Vercel platforms. There is no Krux-managed per-tenant key scheme today.
At a high level: content is ingested and isolated per org by Postgres RLS, retrieved, screened by the guardrails that are live today, and answered by Anthropic with span-level citations and a hash-chained Vault audit record. A data-flow diagram is available under NDA. Contact
security@krux.bio.
Section 4
Data handling.
Residency, encryption, retention, deletion. The procurement-checklist answers.
Runs in a single region today (United States). Per-tenant region pinning is not implemented yet. If you need EU residency, contact security@krux.bio before contracting.
Encrypted at rest by the Supabase and Vercel platforms (AWS underneath). Krux does not run a separate per-tenant key scheme; there is no BYOK today.
TLS for all traffic to krux.bio and trust.krux.bio.
Deleting an RFP or a knowledge file removes the database row and the stored object. Vaults and documents use a soft-delete marker. A scheduled hard-delete after a fixed grace window, and a written retention period, are still being built.
Sensitive content detection
The PHI guardrail flags patient-identifiable shapes (SSN, MRN, and similar) in inputs and outputs. It is rule-based and runs in-process. Configurable per org in the Guardrails admin.
Email security@krux.bio to request deletion. The row-and-object delete path above is what runs today. A documented DSAR procedure with a fixed SLA is in progress.
Section 5
Model providers.
The two providers Krux sends customer text to, and what each receives. Zero-retention and no-training terms are being confirmed at the account level with both providers before we state them here.
Provider
Role
Retention term
BAA
Anthropic
Prompts and retrieved context at inference time. Called over Anthropic's commercial API.
LLM inference (Claude family)
ConfirmingIn progressVoyage AI
Document chunks at embed time, query plus candidate passages at rerank time.
Embeddings and reranking
Confirming—The full sub-processor list, with the infrastructure providers that host the data path, is at
trust.krux.bio/subprocessors. We give 30 days notice before a sub-processor change takes effect.
Section 6
AI governance.
The principles Krux ships on, and the supervisor models that enforce them.
Principles
- 01Citations are the default. A drafted claim with no resolvable source span is stripped before you see it.
- 02Refusal is a feature. Krux flags off-label claims and unsourced clinical claims and routes them to review rather than shipping them.
- 03We do not train any model on your data. We are confirming the no-training and zero-retention terms with Anthropic and Voyage at the account level before we state them as contractual on this page.
- 04Compliance owners configure the guardrails. The product enforces the ones that are live and records that it ran.
- 05Vault activity is hash-chained, and every Vault audit export is Ed25519-signed and independently verifiable.
Supervisor models
Drafted claims
Strips claims that do not resolve to a source span in the knowledge base before the draft is shown.
Drafted claims
Flags unqualified efficacy or safety claims about a named drug or product. LLM-based classifier; findings route to review.
Inputs and outputs
Rule-based detection of patient-identifiable shapes (SSN, MRN, and similar). Runs in-process, deterministic, audit-replayable.
Competitive-IP validator
Roadmap v2Configurable per org
Flags references to competitor confidential product information.
MLR-readiness scorer
Roadmap v2Response drafts
Scores a draft on whether it is ready to route to MLR review.
Section 7
Access controls.
Per-Vault roles that are enforced today, how users sign in, and the source-system connector roadmap. Sign-in is a one-time email code (no password to phish). SAML/SSO and IP allowlisting are not built yet.
Vault roles
Role
Read
Edit draft
Add docs
Members
Guardrails
Export audit
Delete vault
Source-system connectors (roadmap)
Today, documents are uploaded into a Vault directly. ACL-aware ingest from the systems below, where a user removed upstream loses access to the indexed content, is on the roadmap and not yet built. We will move a connector to live here only once it ships.
ACL-aware ingest with re-check at query time. Not built yet.
Microsoft SharePoint
RoadmapSite and folder permission inheritance. Not built yet.
Folder and collaborator ACLs. Not built yet.
Workgroup and ACL inheritance. Not built yet.
Section 8
Operations.
Pen testing, vulnerability disclosure, incident response, sub-processors, and the audit bundle verifier. Stated as it is today, including what is not yet in place.
Penetration testing
A third-party penetration test has not been run yet. We will publish the vendor and a summary once one is scheduled and complete. We will not claim a test that has not happened.
Vulnerability disclosure
Report a security issue to security@krux.bio. We read it and respond. A written triage process with stated SLAs is being put in place.
security@krux.bio ↗Bug bounty
No formal bug bounty program yet. Until one launches, send disclosures to security@krux.bio.
Incident response
Our intent is to notify affected customers promptly and follow up with a written summary. A formal incident response policy with fixed notification timelines is in progress; we will state the timelines once it is signed off.
Sub-processors
The current sub-processor list is at trust.krux.bio/subprocessors. We give 30 days notice before a sub-processor change takes effect.
View sub-processors ↗Audit bundle verifier
A Vault audit export can be verified in your browser at trust.krux.bio/verify against the Ed25519 public keys at trust.krux.bio/keys. No Krux server is involved in the check.
Open the verifier ↗Section 9
Legal artifacts.
The signable documents your procurement and InfoSec teams need before a deal closes.
One contact for everything legal.
DPAs, BAAs, vendor security questionnaires, procurement docs.
privacy@krux.bio