Enterprise Security

End-to-end trust, by default.

Identity, encryption in transit and at rest, secrets management, audit, and hardened defaults. Not a roadmap — this is what ships in the platform today.

OIDC / SAML SSO + SCIMmTLS + certificate rotationExternal vaults (HashiCorp / AWS / Azure / GCP)Audit streamed to your SIEM

Identity & access

Standards-first SSO, fail-closed RBAC, hard isolation.

Single sign-on, standards-first — OpenID Connect with PKCE, discovery and JWKS validation; SAML 2.0 with real XML-signature verification against your pinned IdP certificate (multiple-reference and algorithm-downgrade attempts are rejected). Users are provisioned just-in-time, and SCIM keeps your directory in sync — joiners, movers and leavers propagate automatically.
Two-factor authentication — TOTP with short-lived partial tokens during the second step.
Hardened sessions — Short-lived access tokens with server-side revocation, refresh-token rotation with family reuse-detection — a stolen refresh token that gets replayed burns the whole family. Playground and device tokens are separately scoped and reaped on expiry.
Role-based access control, fail-closed — ADMIN / MEMBER / VIEWER roles enforced at the API layer; when in doubt, the answer is no. Tool access in the MCP layer adds pattern rules, namespace ACLs and pluggable role providers (file, JDBC, LDAP).
Hard tenant isolation — Every organization gets its own scoped runtime — plugins, topics, data and quotas are partitioned per org, and the isolation is enforced by dedicated cross-org test suites, not convention.
Password storage done right — when you use passwords at all — Local accounts are optional and off by default (SSO-first). When enabled, passwords are stored as PBKDF2-HMAC-SHA256 verifiers at 600,000 iterations (the OWASP-recommended work factor) with per-user random salts, constant-time comparison, and version-tagged hashes so the work factor can rise over time — older hashes are transparently upgraded on the owner's next login.

Encryption in transit

TLS and mTLS everywhere it matters.

TLS and mutual TLS across the cluster — Node-to-node replication (Raft / gRPC) supports server TLS, mTLS with client-certificate identity, SPKI certificate pinning, and automated certificate rotation — cluster identity can come from the client cert itself.
Authenticated Kafka front-door — The Kafka-compatible listener speaks SASL PLAIN, SCRAM-SHA-256/512 and OAUTHBEARER; with SCRAM the server stores only salted verifiers — never the password.
TLS on every connector that leaves the box — MQTT, AMQP/RabbitMQ, Redis, Elasticsearch, HTTP and Postgres/JDBC SSL.

Encryption at rest

Focused where it counts, honest about the rest.

Credentials never touch disk in plaintext — Connector and integration secrets are sealed with AES-256-GCM in a dedicated credential vault, with key-id prefixes so keys can rotate without re-encrypting the world.
Encrypted cold storage — Tiered-storage segments are encrypted client-side (AES-256-GCM) before they leave the node; cluster snapshots are encrypted too.
Field-level encryption with data classification — For selectively sealing sensitive columns inside event payloads.

Secrets management

Bring your own vault, fail-hard, masked everywhere.

Bring your own vault — A pluggable secret-provider layer resolves secrets from HashiCorp Vault, AWS Secrets Manager, Azure Key Vault, GCP Secret Manager, environment, or local files — with caching and a uniform ${secret:NAME} syntax in pipeline config.
Deploys fail hard on unresolved secrets — Interpolation (${secret:…}, ${env:…}) refuses to ship half-resolved configuration — a missing secret stops the deploy, it doesn't silently ship a literal placeholder.
Masked everywhere — Secret fields are masked in the UI, exports and CLI output; the local CLI store is permission-locked (0600); traces and trajectories pass through a PII redactor (emails, phones, cards, API keys) before anything is persisted.

Audit & traceability

Who did what, when — and a time machine per event.

A durable audit log — Who did what, to which resource, with what outcome — queryable from the UI and CLI, and streamable live to your SIEM over SSE with scoped stream tokens.
Every AI tool call is audited — In the MCP layer (caller, tool, arguments, outcome, latency), with pluggable persistence.
Per-event time travel — Distributed W3C/OTel tracing plus recorded event trajectories let you replay exactly what a pipeline did — with PII redacted before storage.
Human-in-the-loop approvals — For sensitive actions: route any tool call through an approval gate (Slack / WhatsApp / email) and keep the decision in the audit trail.

Hardened by default

Private by default, abuse-resistant, sandboxed user code.

Private by default — The server binds to localhost until you say otherwise; containers run as a non-root user.
SSRF protection built in — Outbound HTTP from pipelines resolves DNS first (rebinding-proof) and refuses private ranges, loopback, link-local and cloud metadata endpoints unless explicitly allowed.
Abuse resistance — Per-role and per-IP token-bucket rate limits, brute-force login throttling, per-org quotas, and ephemeral sandboxes that are reaped and revoked on expiry.
Injection-hardened storage — Every SQL identifier is validated and quoted; every value goes through prepared statements.
Sandboxed user code — Custom functions run as WASM in a memory-limited, pure-JVM interpreter — not as arbitrary processes on the host.
Signed licensing, fail-closed — License tokens are Ed25519-signed by the vendor and only verified (never signed) inside the product; backup manifests are signature-verified too.
Split-brain defense — Epoch-fenced replication prevents a deposed leader from writing stale data.

The same engine at every tier.

None of this is an "enterprise edition" bolt-on: the free, self-hosted Pulse runs the same hardened engine. Enterprise adds scale, SSO/SCIM at organization level, and support.

Questions about a specific control — pen-test reports, deployment reviews, or compliance mapping?