Skip to content

Security

  • Change the generated administrator password immediately and terminate TLS at a trusted ingress or reverse proxy.
  • Keep SSRF blocking enabled. It is re-evaluated after DNS resolution and on every redirect.
  • Keep robots enforcement enabled unless you own or are explicitly authorized to crawl the target.
  • Run browser workers with minimal Kubernetes permissions, no service-account token, constrained egress, and separate resource limits.
  • Artifact keys are normalized and cannot escape the configured filesystem root.
  • Queue messages contain references and identifiers, never raw content or credentials.
  • S3, database, RabbitMQ, and OpenSearch credentials come from environment/secret stores. Git HTTPS tokens currently live in Git ingestion-job configuration JSON; they are redacted externally but not encrypted at rest, so protect database access and backups.
  • Embedding API keys must be injected from a secret store. Remote embedding providers receive the indexed/query text, so use TLS and approve the endpoint as a data processor.
  • Answer-generation API keys may come from a secret store or the authenticated Settings page. Settings-entered keys are database-resident overrides, so restrict administrative access and protect database backups with encryption. Answer providers receive the question, caller-supplied history, and retrieved chunk context; ContextCrate treats that context as untrusted and does not persist it in audit logs.

Local users are BCrypt-backed through Spring Security. The initial user is created only when the user table is empty. OIDC dependencies are present for deployment integration; provider registration is configured through standard Spring Security properties.

Global TLS trust-all flag

CONTEXTCRATE_TLS_TRUST_ALL_CERTIFICATES disables all TLS certificate and hostname validation for every outbound connection the process makes, removing man-in-the-middle protection process-wide — not just for the endpoint you intend it for. Prefer mounting your internal CA into the JVM truststore (-Djavax.net.ssl.trustStore/trustStorePassword, or add it to the container image's system trust store) so only that CA is trusted, rather than disabling validation entirely. Use the global flag only as a last resort in trusted, isolated environments; the narrower per-job Git/crawler and OIDC trust-all flags are usually the better fit when only one integration needs it.