Resources · Security & trust
Trust is something
you give us evidence for.
AI Warden was built for organisations whose security team has a long memory and a longer questionnaire. This page is the short answer; the long answer ships with the product, signed.
Security posture
Defaults that fail closed.
Read-only by default
The portal opens in read-only. Mutating actions require an explicit role grant. New tenants see no production data without an explicit invite.
Identity-rooted
No standing local accounts on production deployments. Every action is attributable to an IdP-resolvable principal — human or system.
Least privilege
Five-role RBAC by default; finer-grained permissions available. Privileged actions need four-eyes approval. Time-boxed grants.
Encryption
TLS 1.3 in transit. AES-256 at rest with customer-managed keys (KMS or HSM). Per-tenant key separation.
Audit by design
Hash-chained audit log with periodic anchoring. Integrity-verified at read time. Five+ SIEM sinks shipped, including S3 / Snowflake for cold archive.
No telemetry to us
Self-hosted deployments don’t phone home. No anonymous metrics, no auto-update, no usage pings. Network egress to AI Warden is opt-in only.
Deployment topologies
Self-host. Or single-tenant SaaS. Your choice.
AI Warden ships as a single Go binary plus a small Node portal. Postgres for state, ClickHouse for audit. Stateless gateway, scale horizontally. Your VPC, your network policy, your KMS, your IdP.
- Self-hosted — Helm chart, Terraform module, plain-VM systemd
- Single-tenant SaaS — one VPC per customer, in your region
- Air-gapped — offline install bundle, signed releases, no telemetry
- Bring-your-own-everything — KMS, IdP, SIEM, observability
Reference deployment · self-host
your-vpc/ ├── aiw-gateway (3+ replicas · stateless) ├── aiw-portal (2 replicas · stateless) ├── postgres (your managed PG · TLS) ├── clickhouse (your managed CH · TLS) └── kms / hsm (your customer-managed keys) egress: only to vendor LLM endpoints ingress: only from your apps identity: federated to your IdP
Data & keys
What data we touch. What we don’t.
SDLC
How the code that runs in your VPC gets built.
- Two-person review on every PR; mainline-only, signed merges
- SAST on every commit (CodeQL · Semgrep · gosec)
- Dependency scanning + Renovate · pinned versions in releases
- Container image signed (cosign) · SBOM published per release
- Reproducible builds where the toolchain allows
- Annual external pen-test — summary available under NDA
- Quarterly third-party red-team against the LLM & MCP firewalls
- Security-only release branch with backports for current major
Coordinated disclosure
Vulnerability disclosure
Email security@aiwarden.com — PGP key on /.well-known/pgp.txt. We acknowledge within one business day, ship a fix or mitigation within 14 days for high-severity, and credit the reporter in the release notes unless asked otherwise. We don’t pursue good-faith research.
Bounty program in private beta — ask if you’d like an invite.
Certifications & assurance
What’s shipping. What’s next.
Bring your security questionnaire.
CAIQ, SIG Lite, custom — we’ll fill it in and we’ll meet your security team. References from financial services and healthcare available under NDA.