Enterprise AI Risk MatrixThreat Vectors · Architectural Controls · Audit Proof

Every risk of enterprise AI — and how Nunc closes it.

Connecting AI agents to live financial systems, member records, and web portals introduces seven distinct failure modes. Policy memos and prompt instructions cannot stop them; deterministic inline architecture does.

Enterprise AI Threat Vectors Stopped by Nunc Guardian & Private CloudNunc Private Cloud · dedicated DB + vector + vault · zero training sweepInbound Risk Vectors• Raw SSNs, EINs & MRNs• Leaked API & PEM keys• Direct jailbreak turnsRuntime & Tool Risks• Poisoned web DOM / email• Unapproved SQL / API writes• Exposed browser cookies✕ Halted at perimeterNunc Guardian (6 Stages)✓ 17 PII types → ⟦PII:TYPE#N⟧✕ Hard-blocks secrets & keys✕ Strips DOM/email injections✓ SQL/GraphQL scope guard• Human gate on every write✓ DNS allowlist & SSRF block✓ SHA-256 ISO 42001 receiptAuto-quarantine on probingIsolated Containers• Governed Browser• Compute Sandbox✓ Vault-injected authSingle-Tenant Stores• Dedicated DB & Vault• 4-Tier Vector Memory✓ Zero WHERE leaksStateless ModelsSees ⟦PII⟧ tokensZero retention✓ 0 training bytesConnected Stack45+ SaaS & ERPsLive warehouses✓ Writes gated
Interactive Failure vs. Defense Inspector

Trace what happens when an attack or mistake hits runtime.

Select an enterprise failure mode below to compare unguarded AI and generic tool routers against Nunc's inline firewall and single-tenant boundary.

Critical · Data Exposure

Regulated PII, PHI & API keys enter third-party model contexts

When employees or tool routers query CRM, payroll, or donor tables directly, raw SSNs, EINs, bank account numbers, and pasted API keys flow straight into external LLM context windows and vendor logs.

Without Nunc (Unguarded AI & Tool Routers)Exposed
Query result  { member: "Claire Sutton", ssn: "412-88-1904", routing: "021000021" }
Pasted input  "Use key sk-live-98a7f... to reconcile disbursements"
Sent to LLM   Cleartext SSN + ABA routing + live API secret exposed
Guardian Stage 2 Synthetic PII Vault & Secret Hard-BlockNeutralized
Guardian S2   Luhn / ABA Mod-10 checksum + 17-category NER active
Secret check  BLOCKED raw API credential prior to model dispatch
To native AI  { member: "⟦PII:PERSON#1⟧", ssn: "⟦PII:SSN#1⟧", routing: "⟦PII:ROUTING#1⟧" }
To ext. MCP   { member: "[REDACTED:PERSON]", ssn: "[REDACTED:SSN]", routing: "[REDACTED:ROUTING]" }
  • Reversible ⟦PII:TYPE#N⟧ tokens preserve exact multi-table joins and cohort math
  • External MCP clients (Claude, ChatGPT, Cursor) receive one-way [REDACTED] masks
  • Raw API keys, PEM private keys, JWTs, and DB URIs fail closed before inference
Explore Nunc Guardian
The Seven Enterprise AI Risks

Architectural controls outside the model's reasoning loop.

You cannot ask an LLM to police itself. Every Nunc safeguard executes deterministically in gateway infrastructure, isolated containers, and dedicated tenant stores.

Risk 01 · Nunc Guardian
PII, PHI & Secret Exposure in Model Contexts
The Issue
Database queries, CRM lookups, and employee prompts routinely contain SSNs, EINs, bank routing numbers, medical IDs, and API keys that leak into external model context windows.
Nunc Architectural Solution

Guardian Stage 2 swaps 17 regulated entity types into deterministic ⟦PII:TYPE#N⟧ tokens for native reasoning, enforces irreversible [REDACTED] masks on external MCP clients, and hard-blocks raw API, cloud, and PEM keys.

Risk 02 · Guardian + Browser
Direct & Indirect Prompt Injection (Web DOMs & Email)
The Issue
Hidden instructions inside untrusted emails, support tickets, PDFs, or scraped web portals hijack an agent mid-turn into running unauthorized queries or exfiltrating data.
Nunc Architectural Solution

Stage 1 normalizes obfuscated Unicode/Base64 smuggling and Stage 3 pairs deterministic structural rules with semantic classification on both user prompts and untrusted tool/browser DOM outputs — backed by automatic session quarantine.

Risk 03 · Nunc Console
Unauthorized System Mutations & Destructive SQL
The Issue
Giving agents broad OAuth tokens or database credentials lets a single hallucination or hijacked turn mutate ERP ledgers, send external emails, submit web forms, or drop tables.
Nunc Architectural Solution

3-tier default-deny PBAC restricts tool access per user, group, and org; Stage 4 blocks SQL tautologies, UNION dumps, and destructive DDL; and every mutating write pauses in Nunc Console for human diff approval.

Risk 04 · Nunc Private Cloud
Multi-Tenant Row-Filter Leaks (WHERE tenant_id = ?)
The Issue
Shared multi-tenant AI clouds store every customer's prompts, files, and embeddings in common tables where one query defect or cache bug exposes cross-organization data.
Nunc Architectural Solution

Nunc Private Cloud provisions a physically separate relational database, private 4-tier vector memory index, AES-GCM encrypted object vault, and 24-hour auto-purged scratch store for every organization.

Risk 05 · Data Studio & Cloud
Model Training-Log Ingestion & RAG Table Sprawl
The Issue
Copying warehouse tables into third-party RAG stores creates stale duplicate attack surfaces, while consumer or misconfigured model routes sweep prompts into training corpora.
Nunc Architectural Solution

Nunc queries Snowflake, BigQuery, and Postgres in place without copying tables, strips cleartext PII prior to inference, and routes calls exclusively over stateless zero-training paths.

Risk 06 · MCP & Sandbox
Exposed Browser Cookies, Laptop Keys & SSRF Egress
The Issue
Local MCP configs leave static production keys on laptops, raw browser tools expose session cookies to LLMs, and unguarded code sandboxes can probe internal networks.
Nunc Architectural Solution

OAuth 2.1 SSO eliminates static laptop keys; Governed Browser and Compute Sandbox inject credentials inside isolated containers; and Stage 6 enforces fail-closed DNS allowlists and SSRF blocks.

Risk 07 · ISO/IEC 42001 Ledger
Shadow AI Sprawl & Missing Compliance Evidence
The Issue
Security and legal teams lack unified visibility into which AI assistant accessed which connector, whether DLP fired, or who authorized a write during SOC 2, HIPAA, or ISO 42001 audits.
Nunc Architectural Solution

Console attributes every call across Studio, MCP, SDK, and Sidecar, runs continuous AI-DSPM posture checks, provides one-click org kill switches, and writes SHA-256 content-hashed ISO/IEC 42001 audit records.

CISO & DPO Readiness
Ready for your security, legal, and compliance review.

Every Core Platform and Enterprise tenant ships with the complete 6-stage Guardian firewall, dedicated single-tenant data stores, OAuth 2.1 authentication, and exportable audit evidence on day one — never as a paid security add-on.

Summary matrix mapping enterprise AI risks to Nunc enforcement layers and compliance frameworks
Risk CategoryEnforcement LayerFailure PostureMapped Frameworks
PII, PHI & Secret LeakageGuardian Stage 2 Vault & ChecksumsTokenize / Redact · Hard-block keysHIPAA · GDPR · PCI-DSS · SOX
Direct & Indirect Prompt HijacksGuardian Stage 1 + Stage 3 + QuarantineFail-closed on obfuscation; deterministic floorISO/IEC 42001 · OWASP LLM01
Unapproved Writes & SQL Exploits3-Tier PBAC + Stage 4 + Human Sign-OffDefault-deny; writes parked until signedSOX · SOC 2 CC6 · ISO 42001
Cross-Tenant Data BleedNunc Private Cloud Single-Tenant StoresFail-closed tenant routing; 0 shared tablesSOC 2 · HIPAA · GDPR Art. 32
Browser & Sandbox EscapesEphemeral Containers + Stage 6 EgressZero raw cookies/keys; SSRF fail-closedISO/IEC 42001 · Zero Trust

Launch your private AI in minutes — without giving up your data.

Start a 14-day isolated Core Platform trial with $0 due today, inspect Guardian live, or speak with our team about a custom deployment.