Security model
Clyvel is designed as a multi-tenant AI operations control plane. Security controls are built around organization-scoped authorization, explicit provider connections, scoped API keys, protected session handling and auditable operational data.
This page describes implemented product controls and operating practices. It does not claim certifications that Clyvel has not obtained.
Identity and access
- Native account authentication with verified-email flows and Google OAuth support.
- Organization membership with owner, admin and member authorization boundaries.
- HTTP-only session cookies with SameSite protections.
- Gateway API keys are stored as cryptographic hashes rather than reusable plaintext secrets.
- Provider credentials are designed to be encrypted at rest using a dedicated encryption key.
- Revocation and regeneration flows are available for gateway credentials.
Tenant isolation and authorization
- Operational records are scoped to an organization identity context.
- Provider connections and gateway traffic are resolved within the authenticated organization.
- Application and agent scopes are validated before gateway traffic is accepted.
- Legacy authentication paths are disabled by default unless explicitly enabled for compatibility.
Application and transport protections
- Production CORS is restricted to Clyvel's public web origins.
- Security headers prevent MIME sniffing and clickjacking and apply a restrictive referrer policy.
- HTTPS is expected for production frontend, API and OAuth callbacks.
- Private application and authentication routes are excluded from search indexing.
- Sensitive local files, environment files, caches and databases are excluded from Vercel deployment artifacts.
Operational safety
- Gateway traffic is validated against the configured agent and approved model.
- Request telemetry captures status, latency, token usage and cost for operational investigation.
- Telemetry persistence health is surfaced so silent logging failures can be detected.
- Copilot is read-only by default; supported mutations use explicit review and revalidation flows.
- Budget, incident and governance controls are organization scoped.
Secrets and customer responsibility
Do not put provider secrets in prompts, support messages, screenshots or public repositories. Rotate credentials immediately if they are exposed. Customers remain responsible for configuring their providers, restricting who can access organization credentials and applying least privilege to connected systems.
Responsible disclosure
If you believe you found a vulnerability, email security@clyvel.com with a clear description, reproduction steps, affected URLs or components and the potential impact. Do not access data that is not yours, degrade the service, use social engineering, perform denial-of-service testing or publish sensitive details before a reasonable remediation period has passed.
Clyvel will make reasonable efforts to acknowledge credible reports, investigate them and coordinate remediation. A bounty or monetary reward is not promised unless a separate program explicitly states otherwise.
Security limitations
No hosted system can guarantee absolute security. Clyvel does not represent that it is certified under SOC 2, ISO 27001, HIPAA, PCI DSS or another framework unless that certification is explicitly published by Clyvel with current evidence.