Security
This page summarizes the controls in Xberg Pro and Enterprise. For where data lives and what leaves a deployment, see Data handling.
Project isolation
Section titled “Project isolation”Every query runs under PostgreSQL row-level security keyed on the project. The API sets the project for the request, and the database refuses rows that belong to any other project. A stolen key for one project cannot read, write or delete another project’s jobs, collections or audit entries. The constraint lives in the database, so it holds even if application code has a bug. Both Pro and Enterprise enforce it.
Authentication
Section titled “Authentication”- API keys are revocable, hashed before lookup, scoped to one project, and carry a
readorwritescope. See Authentication. - Single sign-on works with any OIDC provider, with email-domain allowlisting.
- Enterprise adds SAML 2.0 and SCIM provisioning.
Outbound request controls
Section titled “Outbound request controls”Outbound HTTP goes through a guard with a deny-by-default IP policy. It applies to webhook delivery, crawling, storage endpoints, OAuth flows and source connectors.
- The guard blocks private, loopback, link-local, multicast and other reserved ranges, and cloud metadata endpoints. It checks the resolved IP, so a hostname that resolves to a blocked address is refused.
- Redirects are re-validated at every hop, up to 5 redirects.
- By default, an outbound request has a connect timeout of 5 seconds, a read timeout of 20 seconds and a total timeout of 30 seconds.
SAFE_HTTP_CONNECT_TIMEOUT_SECS,SAFE_HTTP_READ_TIMEOUT_SECSandSAFE_HTTP_TOTAL_TIMEOUT_SECSoverride them, and the read timeout cannot exceed the total. - Webhook URLs are validated at registration and again on every delivery attempt. A URL that resolves to a blocked address fails permanently, without retries, with
errorset towebhook_url_denied_by_policy(orwebhook_url_not_on_allowlistwhen an allowlist is configured and the host is not on it). - Custom storage endpoints are disabled unless you list them in
STORAGE_ENDPOINT_ALLOWLIST, a comma-separated list of exact hosts or leading-dot suffixes. - Caller-supplied model endpoints are validated at submission and again before each engine call. When
PRO_VLM_ENDPOINT_ALLOWLIST(Pro) orVLM_ENDPOINT_ALLOWLIST(Enterprise) is set, it becomes a closed list: every model endpoint, public or private, must appear on it.
Add network-level egress rules as an independent boundary. For Kubernetes, a NetworkPolicy that admits only the destinations you expect means a flaw in application code still meets a second barrier.
Credentials at rest
Section titled “Credentials at rest”Connected-source credentials and OAuth tokens are encrypted at rest with AES-256-GCM envelope encryption. The key encryption key comes from INTEGRATION_KEK, and an operator can rotate it, which re-wraps stored credentials in the background.
Audit log
Section titled “Audit log”State-changing operations are written to an append-only, hash-chained audit log: project, API-key, integration, crawl and job-submission events. A verification routine proves the chain is unaltered. See Audit log and Audit chain.
Licensing and telemetry
Section titled “Licensing and telemetry”License verification is offline. See Pricing and licensing.
Certifications
Section titled “Certifications”Xberg does not hold SOC 2 or ISO 27001 certification. A self-hosted deployment runs inside your own certified boundary and inherits your controls.
Report a vulnerability
Section titled “Report a vulnerability”Do not open a public issue. See Support.