CSH AI Guard: AI agent governance | Cyber Security House

Architecture & Capabilities

Built for teams who need proof, not promises

Agent & MCP registry

Inventory every agent, model, owner and MCP server in one place.

Tool allowlists

Define exactly which tools each agent may call, per environment.

Prompt-injection defence

Detect indirect injection in user input and retrieved content.

Secret & PII redaction

API keys, tokens and personal data are removed before inference.

Human approval

Writes, payments and shell commands are routed for step-up approval.

Decision log

Every allow, block and approval is recorded with its rationale.

Visual Architecture

Pre-Inference MCP & Agent Firewall

Inspects every Model Context Protocol (MCP) tool invocation, enforcing per-environment allowlists and blocking indirect prompt-injection payloads before execution.

Visual Architecture

DLP Secret Scrubbing & Audit Ledger

Automatically redacts bearer tokens, API keys and personal data before model inference and writes every allow/block decision to an append-only ledger.

Verified Architecture Specifications · E-E-A-T Grounding Reference

Deterministic Technical Controls & Machine-Verifiable Evidence

OpenAPI 3.1 Contract Trust & PQC Center

1. Detection Engine (CSH Sentinel): Git-versioned Sigma rules (CSH-RULE-001..004) mapped to MITRE ATT&CK (T1110.001, T1098.003, T1048, T1190) with idempotent dedupe_key ingestion and SHA-256 immutable event hashes.
2. AI & MCP Firewall (CSH AI Guard): Pre-inference DLP scrubbing ([REDACTED_SECRET], [REDACTED_PII]), explicit per-agent Model Context Protocol (MCP) tool allowlists, and indirect prompt-injection blocking.
3. Database Isolation (PostgreSQL 16): Enforces ENABLE / FORCE ROW LEVEL SECURITY across all 20 tenant tables bound to transaction-local current_tenant_id(); cross-tenant queries fail closed with HTTP 403.
4. Human-Gated Containment: Zero autonomous destructive actions. Host isolation (isolate_host) and credential revocation require named human step-up approval recorded in an append-only audit trail.
5. Verified Exposure & GRC (CSH Scan): Cryptographic DNS TXT or HTTP ownership token required before scanning. Continuous evidence aligned with UK NCSC Cyber Essentials Plus, UK GDPR & Data Protection Act 2018, ISO/IEC 27001:2022.
6. Post-Quantum Roadmap (CSH Quantum): Cryptographic inventory tracking RSA/ECDSA/TLS keys against UK NCSC 2028 discovery, 2031 priority migration (ML-KEM / ML-DSA), and 2035 full PQC completion.

Operated by ARYA AI TECHNOLOGY LTD (Registered in England & Wales, Company No. 17372232) · Canonical Domain: cybersecurityhouse.com Regional Data Residency: UK-LON · EU-FRA · MENA-RUH · APAC-TYO · RFC 9116: /.well-known/security.txt

Defense-in-Depth Isolation

Database-enforced tenant boundaries & UK NCSC PQC alignment.

PostgreSQL 16 FORCE ROW LEVEL SECURITY on all 20 tenant tables.
Pre-inference DLP secret and PII scrubbing before any advisory call reaches ARYA SAI.
Aligned with UK NCSC 2028 / 2031 / 2035 Post-Quantum Cryptography milestones.

Inspect OpenAPI 3.1 Spec

db/migrations/0001_initial_enterprise_schema.sqlPostgreSQL 16

ALTER TABLE security_events ENABLE ROW LEVEL SECURITY;
ALTER TABLE security_events FORCE ROW LEVEL SECURITY;

CREATE POLICY tenant_isolation_policy ON security_events
  FOR ALL TO csh_app_role
  USING (organization_id = current_tenant_id())
  WITH CHECK (organization_id = current_tenant_id());

Isolation: Enforced · Cross-tenant queries return HTTP 403

Frequently asked questions

Does the AI ever take action on its own?

No. Detection is driven by deterministic rules. ARYA SAI only produces advisory summaries from redacted evidence, and every containment action requires a named human approver.

Where is our data stored?

You choose UK (London) or EU (Frankfurt) residency at onboarding. Additional regions are available on Enterprise plans.

Are you certified for ISO 27001 or Cyber Essentials?

CSH Compliance helps you prepare evidence for these schemes. We do not claim certifications we have not been awarded; our own certification status is published on the Trust page.

Can we scan any domain or IP address?

Only assets you have verified ownership of via DNS or HTTP token. Unverified assets cannot be scanned or ingested.

Do you work with managed service providers?

Yes. The partner edition supports delegated tenants, pooled licensing and partner billing.

See it running on your own environment

A 30-minute scoped walkthrough with a security engineer. No slides.

UK Registered (Company No. 17372232)
UK & EU data residency
Available in 15 languages

CSH AI Guard — AI agent governance