Docs/Configure Guardian Firewall
Nunc GuardianSensitivity Tiers · PII Tokenization · PBAC · Egress Allowlists

Configure Guardian Firewall

Configure Nunc's inline 6-stage AI firewall in Console → Trust → Security. Models never hold standing credentials and never receive raw regulated PII unless explicitly permitted by your tenant policy.

1. Step-by-step configuration in Console

  1. Select a Sensitivity Tier (Trust → Security → Data Privacy): Choose Strict (regulated HIPAA/PCI/Legal — masks all 17 PII/PHI detectors and gates all writes), Medium (enterprise default — masks SSN, payment cards, MRN, credentials, emails, and phones), or Permissive (blocks secrets and jailbreaks in internal engineering workspaces).
  2. Toggle PII detectors & add custom regex: Enable or disable individual built-in categories and register custom regex patterns (for example, internal patient MRN formats, claim numbers, or deal codes).
  3. Configure Multi-Tier DLP & Custom Compliance Rules (Data Privacy → DLP & Regulatory Packs): Apply vertical presets or toggle any of the 16 Tier 1 Industry (PCI-DSS, HIPAA, SOX, GLBA, FERPA, COPPA, CUI / 800-171, Legal/IP) and Tier 2 Regional (GDPR, CCPA/CPRA, PIPEDA, UK GDPR, LGPD, APPI/PDPA, FINRA MNPI) packs, add custom { Name, Description, Action } rules (`Block Turn`, `Tokenize / Redact`, or `Audit Warn`), and verify rules live in the Inline DLP Policy Tester.
  4. Set outbound egress allowlists (Trust → Security → Site Permissions): Add approved corporate domains for webhooks, web-fetch, and governed browser sessions. Guardian blocks unlisted outbound HTTP destinations and hidden markdown image beacons.
  5. Configure write-approval channels: Route 409 Awaiting Approval gates to Console approvers or a dedicated Slack/Teams approval channel.
Nunc Console Guardian Data Privacy configuration screen
Console (Trust › Security › Data Privacy): select sensitivity preset, toggle built-in PII/PHI detectors, and configure custom tenant regex rules.

2. How PII masking differs by surface

  • Reversible Tokenization (Data Studio Web): Models reason over deterministic placeholders such as ⟦PII:EMAIL#1⟧ or ⟦PII:SSN#1⟧. Nunc re-hydrates cleartext values only inside the authorized user's browser session — external foundation model providers never receive raw PII.
  • One-Way Redaction (External MCP Clients): When Claude Desktop, ChatGPT, Gemini, or Cursor calls /mcp or /mcp/analytics, Guardian emits irreversible [REDACTED:TYPE] masks so raw PII never enters third-party AI vendor logs.

3. Reference: The 6-stage Guardian pipeline

StageRuntime Enforcement
1. Input InspectionScans user prompts, channel messages, and retrieved chunks for prompt injection, jailbreak heuristics, and restricted topics before model inference.
2. Privacy & Secret TokenizationReplaces PII with reversible session tokens (⟦PII:TYPE#N⟧) in Data Studio, redacts PII as [REDACTED:TYPE] on external MCP responses, and blocks leaked credentials.
3. Tool Policy & Approval GateEvaluates default-deny PBAC (User > Group > Organization), enforces egress domain allowlists, and pauses mutating tools for SHA-256 payload-locked human approval.
4. Output & Egress SanitizationInspects tool payloads, SQL ASTs, and completions to block exfiltration, markdown image beacons, and unapproved outbound URLs.
5. Grounding & Citation CheckVerifies numerical and factual claims against retrieved warehouse rows, tool outputs, and cited tenant documents.
6. Immutable Ledger & TelemetryWrites tamper-evident execution, policy, and approval records to your tenant audit log and surfaces posture alerts in Console.
Nunc Console Guardian Security posture and runtime firewall overview
Console (Trust › Security): inspect real-time prompt injection blocks, PII mask counts, and emergency surface kill switches.

4. Physical tenant isolation & encryption

  • Dedicated physical stores per tenant: Every organization receives its own isolated relational database, dedicated vector index, and separate encrypted artifact namespace — never commingled in shared multi-tenant tables.
  • Zero training on customer data: Prompts, schemas, query results, and documents are never used to train foundation models.
  • TLS 1.3 & envelope encryption: All client, MCP, and Sidecar traffic requires TLS 1.3; connector tokens are encrypted with per-tenant keys.

Related: Guardian Overview · Tool Policies & PBAC · ISO/IEC 42001 AIMS Controls