Friday, November 20, 2026

 





Core Architectural Layers & Responsibilities

1. Consumption & Agent Tier (Untrusted Reasoning Zone)

  • NLP / Agentic Clients: Conversational interfaces, Text-to-SQL planners, or multi-step reasoning agents.

  • Zero-Trust Assumption: Agents operate with no direct network path or static credentials to SQL databases or internal MCP endpoints. Every session carries a short-lived user/service identity token (via Equitus ICAM or enterprise IdP) asserting role, clearance/tenancy, purpose of use, and environment classification.










2. Fusion Secure Context Plane (SCP)

The SCP acts as both an MCP Gateway/Proxy and a Semantic & Policy Control Plane:

  • Identity & Context Binding (ICAM Integration):

    • Extracts user/agent attributes (Role, Org Unit, Security Clearance, Region/Data Residency) and binds them to an immutable session_context_id.

    • Prevents "confused deputy" attacks where a low-privilege user prompts a high-privilege agent into querying restricted tables.

  • Semantic Grounding & Context Assembly (Fusion + KGNN):

    • Maps natural-language business/mission terms to canonical entities and approved relationships rather than exposing raw physical table names.

    • Dynamically constructs a least-privilege context window: only schemas, semantic views, and MCP tool definitions permitted for that user's attributes are injected into the model's prompt/tool manifest.

  • Arcxa Governance, Risk & Lineage Engine:

    • Dynamic MCP Tool Filtering: Intercepts MCP tools/list and tools/call requests, stripping unauthorized tools and validating tool arguments against JSON-schema and policy rules.

    • Deterministic SQL Guardrails (AST Inspection & Rewrite): Parses agent-generated SQL into an Abstract Syntax Tree (AST) before execution:

      • Enforces read-only (SELECT) or approved stored-procedure execution.

      • Blocks SELECT *, cross-tenant joins, Cartesian products, and high-cost unindexed scans.

      • Injects mandatory row-level predicates (e.g., WHERE tenant_id = :ctx_tenant AND classification <= :ctx_clearance) and column-masking/redaction rules (PII, PCI, CUI) via Arcxa compliance tags.

    • Provenance & Lineage Ledger: Attaches Arcxa lineage metadata (query_id, policy version, source tables/documents, row counts, redactions applied) to every response payload so downstream answers remain fully auditable.




3. Execution & Data Plane (Enforced Isolation Zone)

  • Guarded MCP Servers: Domain-specific MCP servers (document stores, REST/ERP wrappers, knowledge graph tools) accept requests exclusively from the SCP mTLS service identity, propagating the signed user context header for downstream auditing.

  • SQL Query Firewall & Unified Views: Databases accept connections only from the SCP SQL Proxy using short-lived, scoped credentials. Queries target curated Fusion views or governed tables rather than raw staging schemas.


Stage

Component

Action Performed

Security / Value Outcome

1. Auth & Context

ICAM + SCP Gate

Validates user token and establishes

session attributes (role, clearance, tenant).

Blocks unauthorized sessions and

binds identity to all downstream hops.

2. Discovery

SCP + Arcxa

Filters available MCP tools and semantic

schema subsets for the active session.

Agent cannot see or invoke tools/tables

outside user policy; cuts prompt bloat.

3. Planning

NLP / LLM Agent

Generates candidate MCP calls and/or

SQL using only the scoped semantic schema.

Lower hallucination rate and

tighter schema accuracy.

4. Enforcement

Arcxa Policy Engine

Parses SQL AST, injects row/column

filters, validates MCP arguments,

checks risk score.

Prevents data exfiltration, SQL

injection, and cross-domain leakage.

5. Execution

SQL Proxy / MCP Server

Executes rewritten parameterized SQL

and approved MCP calls;

masks sensitive fields.

Deterministic, bounded performance

and strict least-privilege data access.





No comments:

Post a Comment

  Core Architectural Layers & Responsibilities 1. Consumption & Agent Tier (Untrusted Reasoning Zone) NLP / Agentic Clients: Conver...