Your AI’s weaknesses stay in your building.
A tool that finds how your AI can be broken holds sensitive findings. SENTINEL-X is self-hosted so those findings, and the keys used to produce them, never leave your infrastructure.
How the platform protects you.
Self-hosted by design
SENTINEL-X runs as one Docker container or on any Node.js host you control. Systems, prompts, replies and results stay on your volume.
Sign-in enforced in production
In production the server refuses to serve without an admin password, because it can reach internal endpoints and uses your API keys.
Secrets stay in your environment
Connected systems and evaluator models refer to keys by environment-variable name. Workspace settings show which ones are missing — never their values.
Scoped automation access
CI and scripts use a dedicated bearer token instead of a user session.
Secure sessions
Session cookies are signed with your own secret and marked secure behind HTTPS.
Retention you control
Set how many days of runs to keep; older runs are pruned automatically. Back up the database with a standard SQLite backup or a volume snapshot.
Every AI decision attributed
Each run records which model played which role, and each human review records who decided, when and why.
Built for operations
A liveness and database health check for your load balancer, and one-container deployment with automatic migrations.
Attack models are kept on a short leash.
The adaptive attacker works only toward the objective of the test it is continuing; it is instructed never to add new harmful goals. Every reply it provokes is still checked in code, and the attack stops the moment an exact check fails.
The rewriter must keep intent, canaries, numbers, emails and names exactly as written. If it changes them, or declines, the original wording runs instead — and the run says so.
Security questions?
Ask the person who built it. Architecture, data flows and deployment options — straight answers.