Skip to main content
Superatom is deployed entirely within the customer’s infrastructure. All platform components run as Docker containers on the customer’s servers. Business data is processed ephemerally and never stored in the platform.

Deployment Model

The entire platform, including the orchestrator, data-source agents, WebSocket server, UI, reporting engine, and audit storage, runs on the customer’s infrastructure. There is no cloud-hosted instance.

Agent Security Model

Each connected data source has a dedicated agent, a containerized service that holds credentials only for its assigned source. The Main Agent (orchestrator) never holds database credentials and never directly accesses customer databases. Each agent runs in its own Docker container with isolated filesystem and memory. A compromised agent cannot access another agent’s data or credentials.

Agent Security Model

Full details on agent isolation, credential separation, and security boundaries

What the LLM Receives

The LLM is a stateless inference service. Each request is constructed from scratch with no memory between calls. Customer data, queries, and knowledge nodes are never used to train or fine-tune models. The LLM provider receives stateless inference requests only.

Authentication

SSO Integration

Superatom delegates authentication to the customer’s existing identity infrastructure:

Session Management

  • JWT tokens with configurable expiration
  • Automatic logout on inactivity
  • All authentication over HTTPS
  • Disabling a user in the IdP immediately revokes Superatom access

Authorization

Role-Based Access Control

Data-Level Permissions

Permissions are enforced at the query planning stage, before any SQL is generated or data is accessed. Rephrasing a question does not bypass permissions.

Encryption


Audit Logging

Every interaction is recorded in an immutable audit trail stored on the customer’s infrastructure: The audit trail is immutable, queryable within Superatom, and exportable to SIEM systems.

Tenant Isolation

For multi-department or multi-business-unit deployments:

SQL Injection Prevention

All queries use parameterized statements:
The AI generates the query structure; parameters are extracted and passed separately. User input is never concatenated into SQL strings.

Prompt Injection Defense

1

Input Sanitization

User inputs cleaned before processing
2

Prompt Structure

System prompts separated from user input
3

Output Validation

AI outputs validated before execution
4

Action Confirmation

Write operations require human approval through configurable approval gates

Company Access

Superatom as a company does not have access to customer deployments. There is no remote access, admin portal, or backdoor. Support is provided using only usage telemetry: query counts, response times, error rates, and feature usage.

Further Reading

Agent Security Model

Agent isolation and credential separation

Data Protection

Zero-storage architecture and encryption

Access Control

Permission system and enforcement

Action Safety

Approval gates and write-back controls

Network & Data Flow

What leaves your network and air-gapped options

Compliance

SOC 2, GDPR, HIPAA, ISO 27001