Zero-Leakage Security Architecture

PII NEVER REACHES
THE MODEL.

RGX intercepts every payload before it hits an LLM — scrubbing PII, PHI, and financial identifiers in real-time. Every workspace is isolated at the data layer. Every redaction is logged to an immutable compliance event trail. Every Workspace Security Token is SHA-256 hashed. No plaintext credentials stored. No payload content persisted. This is what zero-leakage means in practice.

For Enterprise Security Teams
CISO TECHNICAL OVERVIEW
Encryption at Rest
AES-256-GCM. All sensitive configuration data, workspace records, and security token hashes stored on encrypted PostgreSQL volumes. Managed encryption keys, not user-controlled.
Encryption in Transit
TLS 1.3 enforced on all inbound connections and outbound LLM provider calls. HTTP connections are rejected — no plaintext path exists. Certificate pinning available for dedicated deployments.
Security Token Management
SHA-256 hashed before any persistence. Plaintext token displayed once at provisioning and never stored. Authentication validates the incoming token's hash against the stored hash. Even a full database dump cannot reveal the token. 48-character hex tokens from a CSPRNG source.
PII / PHI Tokenization
Automated in-memory scrubbing on every protected request. Identifiers replaced with sequentially numbered tokens ({{REDACTED_PHI_1}}, {{REDACTED_FIN_2}}) before any LLM call or database write. Token map is never stored — raw values discarded immediately after the scrub pass.
Name Tokenization
Person names are tokenized separately from the PII/PHI categories above, for every industry. When RGX generates the response, the real name is restored before it reaches you, so the reply can address the right person by name — the AI model itself never sees it. In passthrough mode (bring your own AI), the name token is not restored by RGX, since there's no return trip through us to reverse it on; you receive it as literal [NAME_N] text. Either way, the name↔token mapping is held only in this process's memory, scoped to the calling session, purged automatically after 30 minutes of session idle time or immediately on session end — never written to disk, logs, or the database.
Zero Raw Text Retention
Payload content (message text, PII, PHI, financial data) is processed in-memory only and is never written to our database, log files, message queues, or any external service. Central logging servers record only metadata: endpoint, timestamp, byte count, token count, status code, and redaction event counts.
Workspace Isolation
Every client workspace is isolated at the database query level — not application-layer logic. A sub-token issued for Workspace A cannot access Workspace B's data. Cross-workspace access returns 403 at the infrastructure level before any query executes.
Immutable Audit Log
Every compliance event (redaction fired) writes a non-blocking entry to a permanent audit table: event_id, workspace_id, tenant_id, seat_id, industry, redaction_count, token_namespace, and created_at. This log cannot be modified or deleted. It is the evidence trail auditors point to.
Infrastructure
Render Services Inc. — SOC 2 Type II certified (Security, Confidentiality, Availability). HIPAA BAA executed with Render. PostgreSQL with encrypted connections. Secrets managed via environment variables — never hardcoded. Production access restricted to authorized RGX personnel with MFA enforced.
LLM Provider BAA
Anthropic, PBC operates under a Data Processing Agreement with RGX Systems that explicitly prohibits use of submitted data for model training. LLM providers receive only tokenized (scrubbed) payloads — raw PHI or PII never reaches them.
Breach Notification
72-hour notification of any impermissible PHI use or disclosure per BAA and 45 CFR §164.410. 60-day notification for Security Incidents per HIPAA Security Rule. RGX will provide sufficient information for covered partners to satisfy their own breach notification obligations.
Dedicated / Air-Gapped
Available for enterprises requiring data residency or VPC isolation. Docker pull or Terraform (AWS ECS Fargate, Azure AKS). Offline RS256 LICENSE_KEY JWT — no outbound RGX network calls. Full deployment reference: /deploy.
BAA → DPA → Data Flow Diagram → security@rgxsystems.com
Every workspace is completely isolated

Each provisioned workspace operates in its own isolated data context. Usage events, billing aggregates, and security token data for one organization are never accessible to another — not through the platform, not through the dashboard, not even to us in normal operations. Complete data segregation is enforced at the database query level.

Security tokens are hashed — never stored

Your Master Organization Token is shown exactly once at provisioning. We immediately SHA-256 hash it before writing anything to the database — the plaintext token is never persisted. Authentication validates the incoming token's hash against the stored hash. Even a full database dump cannot reveal your token.

Payload content is not persisted

Message text and payload content transmitted through the protection layer is processed in memory only — it is never written to our database. We log metadata only: endpoint, timestamp, byte count, token count, and redaction event count. Your clients' regulated data does not live in our storage.

AES-256-GCM at rest — TLS 1.3 in transit

All sensitive configuration data — credentials, tokens, and internal keys used to operate the infrastructure — is encrypted with AES-256-GCM before being written to disk. All data is transmitted exclusively over TLS 1.3. Your data is unreadable without the encryption key even if someone accessed the server directly.

Your payload data is never used for model training

Data that passes through the protection layer is forwarded to our processing pipeline provider under a data processing agreement that prohibits use of submitted data for model training. If you enable passthrough mode to run your own model instead, RGX never forwards your data to any LLM at all — PHI/PII and person names are still scrubbed and tokenized in-memory exactly as above, and the clean text is returned directly to you. Nothing is stored either way.

Rate limiting and abuse prevention

All traffic is rate-limited to prevent abuse and brute-force attacks against your workspace. Invalid token attempts are tracked. Workspaces exhibiting anomalous usage patterns are flagged for review. Your token cannot be guessed — it is a 48-character hex string generated from a cryptographically secure random source.

Built on trusted infrastructure

Hosted on Render (SOC 2 Type II certified). Database on PostgreSQL with encrypted connections. All secrets managed via environment variables — never hardcoded. Infrastructure access restricted to authorized RGX Systems personnel only.

Compliance & Certifications
ENTERPRISE-GRADE COMPLIANCE
SOC 2 TYPE II

RGX runs on Render's SOC 2 Type II certified infrastructure — independently audited for Security, Confidentiality, and Availability. Audit period: October 2024 – September 2025.

GDPR COMPLIANT

Our infrastructure provider maintains a GDPR Data Processing Agreement (DPA). Partner operator data is handled in accordance with GDPR requirements. See our DPA for full details.

WHAT WE WILL NEVER DO
  • ✗ Forward raw PII, PHI, or financial identifiers to any LLM
  • ✗ Store plaintext security tokens — only SHA-256 hashes
  • ✗ Persist payload content (message text, PII, PHI) to the database
  • ✗ Use payload data for model training or profiling
  • ✗ Allow cross-workspace data access — enforced at the data layer
  • ✗ Log AI response content to persistent storage
  • ✗ Access your workspace data without authorization
  • ✗ Contact RGX servers in dedicated/air-gapped deployments
Compliance Engine
THE SCRUB PIPELINE
🛡
Real-time PII/PHI interception

Every protected request passes through the Compliance Engine before any LLM call is made. Text is scanned for PII/PHI using industry-specific pattern sets, identifiers are replaced with numbered tokens ({{REDACTED_PHI_1}}, {{REDACTED_FIN_2}}), and the scrubbed text is what the model receives. Raw identifiers never leave your network.

📋
Immutable compliance event log

Every request that triggers a redaction fires a non-blocking write to a permanent audit table — recording the timestamp, tenant_id, seat_id, redaction count, and pattern types matched. This log cannot be modified or deleted. It is the evidence trail a CISO or auditor can point to.

QUESTIONS ABOUT SECURITY?

Email us at security@rgxsystems.com — we respond to every inquiry.

Get Started Free → Read our Data Processing Agreement →