# Security questionnaire answer bank (SIG-2026-aligned) > Reusable, pre-verified vendor answers for third-party-risk questionnaires — > Shared Assessments **SIG 2026**, CSA **AI-CAIQ**, and ad-hoc buyer > questionnaires. Every answer is the honest current state on **2026-06-12** with a > pointer to the verifiable artifact. Answers that will change at first release are > flagged `[pre-release]`. > > **SIG licensing note (verified 2026-06-12):** the SIG questionnaire is paid, > licensed content of Shared Assessments (corporate license ~US$7,000/yr); a > completed SIG may be delivered to specific authorized recipients but the > questionnaire itself may not be republished. This bank therefore contains **our > answers in our own words**, organized so they transcribe directly into the SIG > 2026 workbook (21 risk domains in 4 control areas, including the *Artificial > Intelligence* domain; the 2026 release added an ISO/IEC 42001 mapping). > Completing an actual SIG workbook requires the subscription — a founder decision, > demand-gated. ## A. Company & product profile | Question theme | Answer | |---|---| | Legal entity | Olivares.AI — sole-proprietor (autónomo) operating under a registered trade name, Spain (EU). `[pre-release: incorporate-on-demand is a deferred decision]` | | Product | Self-hosted control plane for governing AI agents: inventory, R/RW access map (permitted-vs-observed drift), governance/HITL, audit ledger, FinOps, evals, guardrails, compliance evidence. | | Delivery model | **Self-hosted software** (single Go binary / container / Helm / air-gap bundle). No SaaS today; a managed offering is design-stage only. We never operate customer production. | | Data we receive from customers | **None in production.** The product runs in the customer estate; telemetry stays there. Support interactions may include customer-redacted diagnostics only. | | Sub-processors of customer data | **None** (no hosted service). LLM providers are the *customer's* vendors; the product inventories them for the customer's TPRM (`GET /v1/m/models/data-governance`). | | Personnel | Single maintainer (disclosed, with mitigations): see [vendor-viability.md](./vendor-viability.md). | ## B. Information protection | Theme | Answer | |---|---| | Encryption in transit | TLS 1.3; mTLS collector↔core; hybrid post-quantum key establishment (X25519MLKEM768) is the **default**. `docs/SEC-G3-CRYPTO-AGILITY-PQC.md` | | Encryption at rest | Operator-provided (LUKS/filesystem/Postgres TDE) and attested per tenant in the compliance module; BYOK/CMEK envelope encryption with AWS/GCP/Azure KMS for key custody; honest catalog note: a gap until attested in a given deployment. | | FIPS | Optional FIPS 140-3 build linking the CMVP-validated Go module (cert **#5247**, Level 1, active, sunset 2031-04-26). The validation covers the crypto module, not the product. `docs/SCP-09-FIPS-STIG.md` | | Data minimization | By construction: the product stores **relations and hashes, never payloads/secrets/PII**; redacted fields are SHA-256-hashed at write. | | Data classification / DLP | Deterministic PII/sensitivity scanning with persistent labels; deny-closed DLP egress gate; append-only enforcement events. | | Data residency | Region-pinned tenants with deny-closed cross-region guard (403 `residency_violation`); residency attestation + egress scan. `docs/MULTI-REGION-RESIDENCY.md` | | Privacy / data subject rights | RTBF runbook with dual-control, legal-hold veto, sealed erasure receipts, honest limits documented (`docs/RIGHT-TO-ERASURE.md`); retention + legal hold + WORM archival (`docs/RECORDS-MANAGEMENT.md`). | ## C. IT operations & business resilience | Theme | Answer | |---|---| | Availability SLO | 99.5%/28d single-node (the honest current tier); 99.9%/28d committed for the HA tier. `docs/17-PRODUCTION-READINESS-SLO.md` | | HA | Active-passive leader election (Postgres advisory lock), automatic failover via `/readyz` drain, shared audit-signing key custody, online expand-contract migrations. | | Backup/DR | Encrypted `.drbundle` backups (operator KEK), verify-on-restore (fails closed on chain/signature mismatch); RPO 15min–6h (cron tiers) down to ≈seconds (Postgres PITR); RTO <15–30min; drill cadence documented. `docs/DR-RUNBOOK.md` | | Capacity | Published measured baseline (≈1.5k durable writes/s single-writer; Postgres 5.2k/7.8k ev/s at 4/16 writers; SQLite→Postgres crossover at 1–2 concurrent writers). `docs/SIZING-AND-CAPACITY.md` | | Change management | Conventional commits + DCO; CI gates (lint/SPDX/boundary/build/test, `govulncheck`, secret scanning); change/deploy ledger in-product; signed releases. | | Vulnerability management | Published CVE remediation SLA (Critical 7d / High 14d / Medium 30d / Low next release; KEV→Critical); reachability-first via `govulncheck`; OpenVEX for non-reachable; weekly rebuild+rescan. `SECURITY.md` | | Penetration testing | **None performed yet** `[pre-release]`; committed program + cadence: [penetration-testing.md](./penetration-testing.md). | ## D. Security incident & threat management | Theme | Answer | |---|---| | Logging & audit | Append-only, hash-chained, Ed25519 per-event-signed ledger; offline verification; 7-year default audit retention; S3 Object-Lock COMPLIANCE-mode archival with chained manifests. | | SIEM integration | Pull export (CEF/LEEF/syslog/OTLP/OCSF) + push forwarder to SIEM/ITSM sinks + posture export for control towers. | | Incident response | Severity model SEV1/2/3 with comms cadence (SEV1 first update ≤15 min); postmortem triggers tied to error budget. `docs/STATUS-AND-INCIDENT-COMMS.md` | | Vulnerability disclosure | `SECURITY.md`: private reporting, safe harbor, acknowledgement ≤3 business days, triage ≤10, coordinated disclosure ≤90; GHSA→OSV advisories; machine-discoverable via RFC 9116 `/.well-known/security.txt`. | | Threat detection | Guardrail/anomaly findings over agent surfaces (OWASP Agentic ASI01–ASI10 catalog); estate kill switch with two-person re-enable. | ## E. Artificial intelligence (the SIG 2026 AI domain / AI-CAIQ Model Security) | Theme | Answer | |---|---| | Does the vendor use AI in the product? | Yes, narrowly: LLM-as-judge in the evals module (customer-configured providers). The control plane's own logic is deterministic; no customer data is sent to any vendor-operated AI. | | AI governance framework | The product *is* an AI-governance layer; the catalog maps 17 frameworks (EU AI Act, NIST AI RMF + GenAI profile, ISO 42001, SOC 2, ISO 27001, GDPR, OWASP Agentic, CISA/Five-Eyes guidance, …) as data with per-framework pins and disclaimers. | | Model inventory & versioning | Model registry + versions; deny-closed signed model admission (Sigstore); deprecation/lifecycle tracking. | | Model/data lineage | Sealed CycloneDX 1.6 AIBOM per model; SPDX 3.0.1 AI-Profile export; model cards (JSON/Markdown) with explicit `not_recorded` for unknown fields. | | Human oversight | Deny-by-default HITL approvals, quorum/SoD dual-control, break-glass that is itself audited and floor-gated, AAL3 step-up for privileged actions. | | AI incident response | Graduated estate/agent kill switch (one-click stop, structural two-person re-enable); guardian loops with anti-spiral dedup + HITL. | | ISO 42001 | Not certified; readiness mapping with two disclosed honest gaps: [iso-42001-readiness.md](./iso-42001-readiness.md). | | Governing Claude in US-government deployments | The product is deployment-surface-aware for the governed provider (verified 2026-07-03 against Anthropic primary docs): FedRAMP-High paths exist via Claude for Government, Amazon Bedrock GovCloud, and Vertex AI Assured Workloads; ITAR only via Bedrock (IL5; Secret-region IL6). Honest limits we surface rather than hide: Claude Code and Cowork are NOT yet in Claude for Government ("on the roadmap this year" per Anthropic), and the Anthropic Compliance API does not cover Bedrock/Vertex-governed deployments — our per-surface catalog (`connectors/claude-api/surfaces.go`) and coverage findings state this instead of implying parity. | > **Founder decision required:** SIG subscription (~US$7,000/yr) only if/when a > buyer demands the official workbook — this bank answers the substance either way.