Unclassified

42 — Security FAQ

Status: CANONICAL · Commercial OS
Owner: Founder (until security hire) — technical truth sourced from docs/doctrine/ + docs/spec/INTEGRATEWISE_CANONICAL_SPEC_v1.0.md
Last verified: 2026-07-21
Honesty rule: this document never claims a certification or control we do not hold. Aspirational items carry explicit status. One overclaim in a security review ends the deal and the reference — accuracy outranks impressiveness.

Purpose

The buyer-facing security answer set — delivered proactively at POV kickoff (arriving prepared is the trust move) and used verbatim in security questionnaires.

Target audience: Persona 4 (CISO/IT), P3 (RevOps evaluating), procurement.

1. Architecture (the five answers that matter most)

"Where does our data live, and who can see it?"
Your tenant's data is isolated with row-level security enforced at the database layer — tenant isolation is a database property, not an application promise. Data is encrypted in transit (TLS) and at rest. Cross-tenant access is structurally impossible through the application path; every request is identity-bound and tenant-scoped server-side. Client-supplied tenant/role claims are never authoritative.

"Can the AI read everything in our tenant?"
No. AI participants receive scoped projections for a declared purpose — never unrestricted tenant state. Every cross-boundary exchange carries an explicit context boundary: subject, source, target, scope, purpose, permissions, policy references, and expiry. No projection may contain credentials, raw secrets, hidden prompts, or another agent's private reasoning.

"Can the AI write to our data or act in our systems?"
Not directly — ever. Two mandatory human gates: (1) memory: AI proposes knowledge; a human approves; only then does it enter organizational memory (TruthLayer); (2) action: execution requires an approved proposal; a prepared plan is never treated as completed work, and an AI message is never treated as an approval. Both gates are architecture, not configuration — there is no admin toggle that removes them.

"What happens when a check fails?"
The platform fails closed. When identity, tenant membership, policy, or boundary verification cannot complete, the request is rejected. A prompt cannot talk the system out of policy.

"What's the audit story?"
Every proposal, approval, execution, sync, and memory promotion is logged with identity attribution and preserved provenance. Evidence, decisions, and authority are recorded separately — you can always answer who knew what, who decided what, on what basis. Audit records are tamper-evident and survive user offboarding.

2. Data Handling

Question	Answer
Model training on our data?	No. Customer data is never used to train foundation models. No cross-tenant data use.
Which AI providers?	Provider-agnostic by architecture; models are swappable by configuration. Enterprise agreements can pin/exclude providers. Truth and memory live in your Spine — a model swap loses nothing.
Connector credentials?	OAuth-based; tokens stored encrypted; connectors read through governed pipelines — no connector writes directly to canonical state. Writebacks are governed operations requiring approval.
Data export / offboarding?	Full Spine export in open formats, anytime. On termination: 60-day retrieval window, then deletion with certification (32 Licensing).
Sub-processors	Cloud infrastructure (edge compute/storage), connector OAuth infrastructure, AI inference providers — current list supplied with the DPA.
Data residency	Multi-region edge infrastructure today; region-pinning on the Enterprise roadmap (status: roadmap — do not sell as present).

3. Compliance Status (dated honesty — update monthly)

Item	Status (2026-07)
SOC 2 Type I	In preparation — controls program underway; target: post-seed funding milestone (68)
SOC 2 Type II	Follows Type I observation window
GDPR	DPA available; data-subject workflows supported; EU residency pending region-pinning
Penetration testing	Scheduled with SOC 2 program; summary shareable under NDA when complete
SSO/SAML, SCIM	Enterprise tier — status: roadmap-near; verify current build before committing dates

Sales rule: deals gated on certifications we lack are nurture, not push (33 Anti-ICP). The honest sentence: "Type I is in preparation with [target window]; here's our architecture packet meanwhile — most teams find the mandatory-approval design answers the risk question the certification approximates."

4. The Shadow-AI Argument (offense, not defense)

Your team already uses AI — surveys say ~80% adopted it, and most use it ungoverned, pasting account data into consumer chatbots with auto-written memory. IntegrateWise is the governed alternative: same capability class, inside your tenant, with scoped projections, approval gates, and audit. The security conversation isn't "new risk?" — it's "replace unmanaged risk with governed capability."

5. Incident Posture

Fail-closed architecture limits blast radius by construction; audit trail supports forensics; contractual breach-notification commitments in the DPA. Public incident-response statement template maintained in 65 Press Kit.

KPIs: security-review pass rate; review cycle time (target < 2 weeks with proactive packet); questionnaire-answer reuse rate (target: 90% from this doc verbatim); zero overclaim incidents — permanent.