Skip to main content
Velor

Security & Trust

Encryption, audit trails, and controls that pass security review on day one.

Your data stays yours: Velor does not use it to train models, and every sub-processor that receives customer content in readable form is named on our sub-processor list, with what it receives and, where known, how it handles retention. Here's exactly how the platform is built — layer by layer.

The documents a security review receives, with their versions, and the things that do not exist yet, such as a SOC 2 report or an independent penetration test report, are listed in the security package on the Trust Center.

Last updated October 2026

Key controls at a glance.

TLS everywhere

TLS 1.2+ on all endpoints including the widget, webhooks, and voice WebSocket (the minimum version is a Cloudflare and Supabase setting that our repository cannot show). Our application sends HSTS on its responses.

AES-256 at rest

All data at rest encrypted with AES-256. Connector credentials are encrypted at the application layer with AES-256-GCM under a key held only on the Velor side, in addition to storage encryption.

Row-level security

Every customer-scoped table our migrations create enforces org membership in a Postgres row-level-security policy. Server code that uses the service role bypasses it by design and relies on an organization filter in each query, which a test suite checks.

HMAC-signed webhooks

Webhook calls signed with HMAC-SHA256 over a timestamp and the body. A replayed inbound delivery has no effect: where the provider timestamps its deliveries we reject anything outside a 5-minute window, and otherwise each delivery is either processed once under a per-delivery idempotency key or applied as an idempotent update, so a replay changes nothing. Outbound deliveries carry the timestamp and an idempotency key so your endpoint can reject stale replays and de-duplicate retries.

No model training

Customer conversations and KB content are never used to train, fine-tune or improve any model that is made available to a third party.

DPA available

Data Processing Agreement (with Standard Contractual Clauses) available on request. Velor acts as your processor under GDPR and a service provider under CCPA.

Audit log

Append-only record of sensitive admin actions — API keys, member changes, SSO config. Owner/admin read. Retained 365 days by default for workspaces created from 31 October 2026 (30 days before that), configurable from 90 to 365 days.

Scoped API keys

Customer API keys carry explicit scopes. SHA-256 hash stored — never the full key. Revocable at any time.

SAML 2.0 / SSO

Built for Okta, Entra ID, Google Workspace and OneLogin, and not yet tested with a customer's identity provider. Domain enforcement. SSO can be made mandatory org-wide.

Infrastructure

Infrastructure.

Compute
The product runs on Cloudflare Workers — edge compute where each request runs in an isolated V8 environment, not a shared VM. Durable Objects manage session-bound state. Every deploy is an immutable Worker version with rollback available.
Database
Supabase-managed Postgres (AWS us-east-1) with row-level security enabled on every customer-scoped table. Supabase publishes a SOC 2 Type II report. Server code that uses the service role carries its own organization filter on each query.
Encryption at rest
All data at rest encrypted with AES-256. Connector credentials are encrypted with a key held on the Velor side before being written to the database — a database dump alone does not reveal them.
Encryption in transit
TLS 1.2+ on every endpoint, including the embedded widget, webhook callbacks, and the voice WebSocket. HSTS is enforced by Cloudflare at the edge.
Data residency
Persisted data is stored in the United States (AWS us-east-1), apart from live voice-call state, which a Cloudflare Durable Object holds for the length of the call; a five-minute cache of knowledge-base retrieval results and Cloudflare request logs sit on Cloudflare's global edge. EU, UK, and Swiss customers are served from US infrastructure under Standard Contractual Clauses (and the UK Addendum where applicable) — these are included in our DPA. EU-resident processing is not available today; if it is a hard requirement for your procurement, tell us early.

Data & Privacy

Data & privacy.

Tenant isolation
Every customer-scoped table carries an organization_id enforced by a Postgres row-level security policy. From a customer-authenticated session, row-level security refuses cross-org reads; server routes that use the service role rely on an organization filter in each query instead.
No model training on customer data
Velor does not use customer conversations, knowledge, or tool payloads to train, fine-tune or improve any model that is or will be made available to a third party. Anthropic's commercial API terms do not permit it to train models on the inputs and outputs we send.
Configurable retention
Data retention is configurable per organization. Conversation transcripts, messages and metadata are retained by default for 365 days after the conversation's last message (or after it was created, if it never had one), whether or not it was resolved; an owner can set any window from 30 to 730 days in Settings → Data & privacy. Action invocation audit logs are purged automatically after 30 days by a daily cron job.
GDPR: export and erasure
Export and erasure requests are supported via the dashboard (Settings → Data & Privacy). An owner can request a full export of conversations, knowledge, and metadata at any time; it is delivered on the DPA's timelines. After contract termination we preserve data for 30 days for export on written request (DPA); an owner who uses Delete organization in the dashboard starts deletion immediately. Deleting an organization removes its database records and makes a best-effort sweep of its stored conversation attachments, avatar, uploaded knowledge files and eval dataset files. Call recordings held at Twilio are deleted through the customer's Twilio credentials where our access still exists, and by us on request otherwise.
Sub-processors listed
Every sub-processor we engage is named at velorplatforms.com/subprocessors, with what each one receives. That includes the providers that receive conversation or knowledge-base text (Anthropic, OpenAI) and those that receive call audio or reply text on the voice channel (Twilio, Deepgram, ElevenLabs).

Authentication & Access Control

Authentication & access control.

SAML 2.0 SSO
Built for Okta, Microsoft Entra ID, Google Workspace, OneLogin, and any SAML 2.0-compliant IdP; not yet run end to end with a customer's identity provider, so we run a joint test with you first. Domain enforcement ensures only corporate email addresses in the configured domain can use SSO login. Configuration is managed by the organization owner. Setup guide →
SSO enforcement
Organizations can require SSO and block password or OAuth login for all members. This prevents credential-based access outside the corporate IdP.
SCIM 2.0
SCIM 2.0 user provisioning and deprovisioning, built for Okta and Microsoft Entra ID. When a user is removed from your IdP, SCIM deactivation signs them out and blocks new sign-ins; an access token already issued can remain valid until it expires. SAML and SCIM have not yet been run end to end with a customer's identity provider; we run a joint test with you before rollout.
MFA
TOTP-based two-factor authentication is available to every user, and enforceable by org policy: when an owner turns on the workspace requirement, the server refuses dashboard and API sessions that have not completed the second factor. Velor's own policy is to require MFA on its administrative accounts (Google Workspace, GitHub, Cloudflare, Supabase, Stripe); an independent check of those accounts is an open item in our SOC 2 preparation and has not yet been performed.
Role-based access
Owner, Admin, Agent, and Member roles each have distinct server-side permission checks. Members cannot open the billing pages or read invoices; agents cannot manage connectors; only owners and admins can invite teammates, and only owners can configure SSO, SCIM, MFA enforcement and the IP allowlist.
Scoped API keys
Customer-facing API keys (velor_live_…) carry explicit scopes. The full key is shown once at creation; only a SHA-256 hash is stored. Keys can be revoked at any time from the dashboard. Each key records when it was last used (updated at most once a minute) for access review.
Audit log
Append-only record of admin actions, which cannot be edited or deleted through the app or API — API key creation and revocation, member changes, SSO configuration, knowledge source changes, webhook configuration — with actor, timestamp, and IP address. Entries are tenant-scoped; only owners and admins can read. Rows are removed only by your configured retention window (365 days by default for workspaces created from 31 October 2026, 30 days before that; an owner can set any window from 90 to 365 days in Settings → Security) or when the organization is deleted.

Application Security

Application security.

Row-level security
Tenant isolation enforced at the database layer on all customer-scoped tables via Postgres policies. Server code that uses the service role bypasses it by design and relies on an organization filter in each query, which a test suite checks.
API authentication
Bearer API keys on every /api/v1 endpoint except the public index and the OpenAPI document. Keys are 256-bit random tokens, stored only as a SHA-256 hash, with per-key scopes, per-key and per-org rate limits, an optional organization IP allowlist, last-used tracking, and immediate revocation.
Widget origin allowlist
Each workspace can restrict which websites may load its chat widget (Settings → Widget → Allowed origins); every chat, consent, attachment, visit-tracking and error-report endpoint then refuses other origins, and requests with no Origin header (the satisfaction-rating endpoint uses a single-use token instead). The list is empty by default, which accepts any website, so existing installs keep working. While it is empty, Widget Settings and the onboarding install step show a warning listing the sites that opened a chat session in the last 7 days, with a one-click option to allow only those.
CSRF protection
Next.js checks the Origin header against the host on server-action requests, and our own guard rejects cross-origin requests to cookie-authenticated API routes. A request that carries no Origin header is not rejected by the framework check.
Input validation
User input is validated on the server: by schema on most write endpoints and field by field elsewhere. A few public actions apply only a length or format check, listed in the public surface matrix we give testers.
Webhook signatures
HMAC-SHA256 signed payloads on webhooks sent to configured endpoints and SIEM destinations (workflow webhook-call steps are not signed). Signature is included in the X-Velor-Signature header. Receiving servers can verify before processing.
SSRF protection
Outbound action calls reject IP literals, RFC1918 private ranges, link-local addresses, and unresolvable hostnames. HTTPS-only enforcement at dispatch.
Secret handling
No known credential shapes are stored in client-side code, and logs redact secret-named fields and credential-shaped values. Production secrets live in Cloudflare's secret store, never in the source bundle.

AI & Data Handling

AI & data handling.

Powered by Anthropic Claude
Language-model inference runs on Velor's own Anthropic API account (OpenAI is used as a backup only if the failover is enabled), under Anthropic's commercial terms, which do not permit Anthropic to train models on the inputs and outputs we send. Retention of those inputs and outputs on Anthropic's side follows the same standard terms. Your conversations do not improve anyone else's product.
Grounded responses
Every turn retrieves from your organization's uploaded content — never the open internet — and the AI is instructed to say it does not know when that content does not cover the question. Replies sent by email, SMS, WhatsApp and Slack are checked before they go out by default (and chat and voice can be switched on in Guardrails): the check looks for specifics not supported by your knowledge base, a held answer is replaced by a safe reply, and your team is notified.
PII redaction on voice
Voice transcripts are pattern-scanned before storage: card numbers, SSNs, US bank routing numbers, IBANs, emails, phone numbers and dates of birth are masked (the last four digits of a card or an email domain may be kept); names and street addresses are not detected. Call recordings (voicemails, warm-transferred calls, and whole calls on numbers where call recording is turned on) are stored by Twilio, not Velor; see Call recordings in our Privacy Policy for when they are deleted.
AI provider retention, per provider
Each AI provider retains data under its own terms and our account settings, and the sub-processor list states each one. Anthropic (inference) and OpenAI (knowledge-base embeddings, and backup inference only while the inference failover is enabled) operate under their standard API terms. Deepgram (speech-to-text) is sent its model-improvement opt-out on every request. Voice numbers run on the customer's own Twilio, Deepgram and ElevenLabs accounts — the dashboard asks for those keys when a number is added — so call audio, transcription and synthesis are retained under the customer's agreements and settings with those providers; ElevenLabs' Zero Retention Mode is opt-in, not a default. Voyage AI retains API data for training by default and is not enabled.

Compliance & Certifications

Compliance & certifications.

SOC 2
Velor does not currently hold a SOC 2 report. Enterprise customers can request our written security questionnaire, review our published controls (above), and our sub-processor list today — no waiting.
GDPR
Velor acts as a processor of your customers' personal data. A Data Processing Agreement is available on request. Email legal@velorplatforms.com with “DPA request” in the subject line.
CCPA
Velor acts as a service provider to its customers. A California resident who chatted with a Velor customer directs access, deletion and opt-out requests to that customer, which controls their data and carries out deletion and export from its own dashboard; we forward a request emailed to privacy@velorplatforms.com to that customer, as the DPA commits. Customers themselves make requests from Settings → Data & Privacy or by email.
Sub-processors
Anthropic (AI inference), OpenAI (embeddings; backup AI inference only while the failover is enabled), Voyage AI (alternative embeddings, not enabled), Supabase (database, SOC 2 Type II), Cloudflare (edge compute and a five-minute retrieval cache), Resend (transactional email), Twilio (voice/SMS), Deepgram (speech-to-text), ElevenLabs (text-to-speech), Stripe (payments), and Sentry where configured. Full list of providers that receive customer content in readable form, with what each receives, at velorplatforms.com/subprocessors.
No PHI or card data
Velor does not currently position itself as HIPAA-compliant or as a PCI data handler. If your use case involves PHI or raw card data, discuss scoping with us before integrating.

Incident Response

Incident response.

Notification commitment
Security incidents that affect customer data are disclosed to affected customers without undue delay and within 72 hours of Velor becoming aware of them, as the DPA commits. Notifications include the nature of the incident, data involved, and remediation steps taken.
Contact
Report a security incident or concern at security@velorplatforms.com. A person reads this inbox on business days.
Responsible disclosure
If you believe you have found a security vulnerability in Velor, email security@velorplatforms.com with “Security report” in the subject line. Give us a reasonable window to investigate and remediate before public disclosure. Valid reports are credited on the disclosure page if the reporter wants it. The full policy — scope, safe harbour, and response times — is at /security/disclosure.

Penetration Testing

Penetration testing.

Third-party penetration testing
Velor has not yet engaged a third-party security firm for a formal penetration test. Enterprise customers may run their own time-boxed test in tester-owned organizations on production, or in a separate environment we build for the engagement, under a written rules-of-engagement agreement — contact us to scope. We will publish the firm and scope here when we begin a formal engagement.
Customer-led testing under NDA
For customers who run their own test, we share our internal architecture overview and our data-flow documents and the relevant Cloudflare configuration under a mutual NDA so the test is informed rather than blind. Contact security@velorplatforms.com.

System status

Availability sampled every 6 minutes, with the measured month-to-date figure published.

Check status page

Send us your security questionnaire.

Request the DPA, ask a specific question, or send the questionnaire directly. A person answers these directly — there is no procurement layer in between.

Last updated October 2026 · security@velorplatforms.com