Security you can
hand to procurement
Agents act with real credentials. We treat that as a security boundary, not a feature. Every control below is in force for all paid plans unless noted.
Documents, on request, without a sales call
Audit reports, penetration test summaries and architecture whitepapers are available under NDA from the trust portal. Security questionnaires are answered in three business days.
Audit reports
SOC 2 Type II and ISO 27001 certificates, current period, with the auditor's scope statement.
Penetration tests
Annual third-party application and network tests, plus continuous automated scanning.
Architecture review
A 22-page design document covering data flows, key management and tenancy isolation.
How the platform protects data
| Area | Control | Detail |
|---|---|---|
| Encryption | TLS 1.3 in transit, AES-256 at rest | Keys rotated quarterly; Enterprise can bring customer-managed keys via KMS. |
| Tenancy | Row-level isolation with per-tenant keys | Vector indexes and object prefixes are namespaced; cross-tenant queries are rejected at the query planner. |
| Access | RBAC plus attribute-based policy | Least privilege by default; production write scopes require explicit grant. |
| Authentication | SAML 2.0, OIDC, SCIM | MFA enforced for Veltrix staff; hardware keys required for production access. |
| Audit | Immutable, append-only log | Every tool call, approval and policy change, exportable to S3, Splunk or Datadog. |
| Residency | US and EU regions | Data stays in the region selected at tenant creation, including backups. |
| Model training | Never | Customer prompts and completions are not used to train any model, ours or a provider's. |
| Retention | Configurable per dataset | 24 hours to 7 years; deletion is hard-delete within 30 days of request. |
| Vulnerability mgmt | Continuous scanning, 30-day SLA | Critical findings remediated within 72 hours; dependency policy blocks known CVEs in builds. |
| Business continuity | Multi-AZ, RPO 5 min, RTO 1 hour | Restore drills quarterly; status history published at status.veltrix.com. |
The threats other platforms ignore
Prompt injection, tool escalation and data exfiltration are the failure modes that actually hurt. Each has a dedicated control, and each is tested in our release pipeline.
- Untrusted content is tagged and cannot issue tool calls without policy consent
- Tool arguments are schema-validated before execution, not after
- Egress allowlists per environment, with DNS and IP pinning
- Secrets are injected at call time and never enter model context
- Canary tokens detect attempts to read across tenant boundaries
Responsible disclosure
Report a vulnerability to [email protected]. We acknowledge within one business day, keep you updated through remediation, and pay bounties from $500 to $25,000 depending on severity. Please do not test against customer tenants.
PGP key fingerprint:
9F2C 41A7 88D1 0B6E 3C55 A1D9 7E04 2B88 5F31 6C0A
Who touches data, and why
Customers are notified 30 days before a new subprocessor is added and may object.
| Subprocessor | Purpose | Location | Data categories |
|---|---|---|---|
| Amazon Web Services | Infrastructure | US, EU | All tenant data |
| Cloudflare | Edge protection, DNS | Global | Request metadata |
| Datadog | Platform telemetry | US | Redacted service logs |
| Stripe | Billing | US | Billing contact, payment tokens |
| Anthropic / OpenAI / Google | Model inference (optional) | US | Prompt content, no training |
| PagerDuty | Incident response | US | On-call metadata |
Need the full security package?
We will send the SOC 2 report, pen test summary, DPA and architecture document in one bundle.