Security & data protection

1. Data encryption

• In transit: all traffic is HTTPS/TLS only. Plain HTTP and requests with embedded credentials are rejected; outbound integration URLs are TLS-enforced.
• At rest: all data sits in Microsoft Azure Storage with Storage Service Encryption, under a customer-managed key (CMK) on the storage account.
• Sensitive credentials (PSA API secrets, Microsoft 365 bot secret, Slack tokens, Twilio tokens, IT-documentation and CIPP keys, SMTP and bring-your-own AI keys) get a second, application-layer envelope: AES-256-GCM with a per-value IV and authentication tag, keyed from a dedicated master key. Encryption fails closed — if the key is unavailable the platform refuses to write plaintext, and it will not start against an unresolved key reference.
• Payment data never touches KYRA servers — card handling is delegated to Stripe (PCI DSS Level 1).

2. Access control & authentication

• Sessions: first-party HttpOnly, Secure, SameSite session cookies; the admin surface is cookie-only. Sessions can be revoked server-side immediately (on sign-out, and automatically whenever a role is changed or removed).
• CSRF: mutating requests use a double-submit token that must match the session.
• Passwords: salted and hashed; sign-in and registration are rate-limited.
• Role-based access control: five built-in roles (global admin, customer admin, support agent, viewer, QA) plus MSP-defined custom roles, enforced centrally by a deny-by-default route-to-area map. Roles can be scoped to specific customers, and no role can grant itself user-management rights (privilege-escalation guard).
• Multi-factor authentication: enforced for production administrative (KYRA staff) access.
• Step-up authentication: particularly sensitive administrative actions can require re-authentication, where enabled by policy.
• Single sign-on: optional Microsoft Entra (Azure AD) SSO, invite-only and matched by verified email.

3. Application security

• Tenant isolation: every record is partitioned by MSP / customer, and customer-scoped routes verify ownership before returning data.
• Injection defence: all datastore filters escape user input; request bodies are validated and typed.
• SSRF protection: any URL the platform will call (PSA endpoints, custom AI endpoints, outbound webhooks) is checked at both save-time and fetch-time — HTTPS-only, no embedded credentials, and private, loopback, link-local, CGNAT and cloud-metadata address ranges are blocked (IPv4 and IPv6), defending against DNS-rebinding.
• Webhook authenticity: supported provider webhooks are signature-verified against the exact request body, with replay controls and durable idempotency for externally retried operations (including billing).
• Telephony safety: emergency, premium-rate, short-code, service-number and unsupported-country destinations are blocked at configuration and call time.
• Safe deploys: a pre-deploy gate blocks release on any syntax error, missing access-control mapping, or duplicate route; changes go through adversarial security review before shipping.

4. Infrastructure, hosting & data residency

• Cloud: Microsoft Azure, UK South region.
• Compute: Azure Functions (serverless) — no customer-managed servers to patch.
• Data at rest: Azure Table and Blob storage in UK South, with Storage Service Encryption and CMK (see §1).
• AI & telephony: AI inference and embeddings run on Azure OpenAI in UK South, and telephony (numbers, calls, SMS) on Azure Communication Services in the UK. Real-time voice (Azure AI Voice Live) runs in the EU (Sweden Central). Support content is not routed to US AI providers by default.
• Segregation: a single logical platform with strict per-tenant data partitioning; no shared credentials across MSP tenants.

5. Data protection, privacy & sub-processors

KYRA processes personal data under UK GDPR as a processor acting on the MSP’s instructions. International transfers are covered by the UK IDTA / EU Standard Contractual Clauses.

The authoritative, dated sub-processor list and the full privacy notice live at /privacy.html.

6. AI & customer-data usage no training

• Not used for training: customer conversations and data are used only to answer the request at hand — they are not used to train foundation models.
• Provider choice: the default provider is Microsoft Azure OpenAI in the UK (UK South); Enterprise MSPs can pin a different provider or region, or bring their own (Azure OpenAI, OpenAI, Anthropic, or a self-hosted Ollama). BYO keys are encrypted and their endpoints SSRF-checked.
• Knowledge-only documentation: the IT-documentation integration ingests knowledge/how-to articles only — it never reads passwords, secrets or configuration assets, and every ingested article is treated as untrusted reference data.
• Grounded & bounded: the technician copilot is grounded strictly in the MSP’s own systems (knowledge base, docs, past tickets), never the open web, and is advisory — it never performs account changes on its own.

7. Business continuity, backup & retention

• Durability: data lives in Azure Storage with Microsoft’s built-in redundancy, blob versioning and soft-delete; deployments are versioned and roll back to the previous package.
• Backup & recovery: required Table and Blob data is exported, encrypted, to a separately-permissioned geo-redundant (GRS) recovery account; restore and deployment-rollback procedures are tested periodically, and backup access is restricted and logged.
• Retention: a configurable, UK-GDPR storage-limitation schedule with platform, per-MSP and per-customer overrides and safety floors — e.g. audit logs and billing usage retained ~6 years, failed-chat capture 180 days.
• Deletion: deleting an MSP archives a financial-only record for the statutory 7 years, then purges the operational data.
• Availability monitoring: a public status page and automated component probes with incident alerting.

8. Logging, monitoring & audit

• Audit trail: privileged and mutating actions (customer edits, role and configuration changes, billing actions) are written to a protected, access-controlled audit log with actor, action and timestamp, retained ~6 years. Audit records are protected and monitored, but are not described as technically immutable unless exported to an appropriately controlled immutable store.
• Telemetry & alerting: application performance, errors and request traces via Azure Application Insights, with alerts for authentication anomalies, application and webhook failures, queue backlog or poison, backup failure and billing anomalies.
• Integration health: the platform proactively probes each customer’s connected PSA and messaging channels and alerts the MSP on failure or recovery.
• Rate limiting & abuse controls: layered in-memory and cross-instance limits on authentication, registration, public chat and the API, with anti-spam guards on end-user channels.

9. Vulnerability & change management

• Least privilege & secrets: platform secrets are held as Azure Key Vault references (with purge protection) rather than in plaintext configuration; stored third-party credentials get an additional AES-256-GCM envelope that fails closed if the key is unavailable.
• Secure SDLC: every change passes an automated pre-deploy gate (syntax, access-control completeness, duplicate-route detection) and unit tests; security-sensitive changes get an adversarial multi-reviewer pass before release.
• Responsible disclosure: report suspected vulnerabilities to the contact below; we aim to acknowledge within two business days.

10. Questionnaire mapping & contact

The sections above map to the core domains of SIG Lite and the CSA CAIQ: encryption (§1), access control / IAM (§2), application & interface security (§3), datacentre & infrastructure (§4), data governance / privacy & supply chain (§5–6), business continuity & data retention (§7), logging & monitoring (§8), and threat & vulnerability management (§9).

For a completed SIG Lite / CAIQ workbook, a signed Data Processing Agreement, or a security review call, contact [email protected].