Skip to content

thecelestialmismatch/HoundShield

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

215 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Hound Shield — AI Compliance Firewall

Hound Shield

Stop your team from leaking CUI to ChatGPT.

A local-only AI proxy that scans every prompt in under 10ms — before it leaves your network. One URL change. CMMC Level 2, HIPAA, and SOC 2 enforced simultaneously.


Live License Latency Models Frameworks Stack Status


News

  • [Apr 2026] Brain AI knowledge base expanded: full CMMC, HIPAA, SOC 2, managed agents, and multi-framework integration coverage.
  • [Mar 2026] Production launch on Vercel. SOC 2 + HIPAA + CMMC Level 2 simultaneously enforced. All 16 detection engines live.
  • [Mar 2026] PDF compliance reports shipped (Growth+ tier). Blockchain-anchored audit trail on Base L2 live.
  • [Feb 2026] Azure Sentinel connector merged. Splunk HEC batch delivery shipped.
  • [Jan 2026] SIEM integration suite complete: Slack Block Kit, Teams Adaptive Cards, Splunk, Sentinel, CEF.
  • [Dec 2025] AES-256 quarantine vault + HITL review workflow shipped.

What Is Hound Shield?

Every time your team uses ChatGPT, Claude, Copilot, or Gemini — they may be leaking API keys, patient records, defense contract numbers, or trade secrets to a third-party AI that stores your data indefinitely.

Hound Shield is a single URL change. It sits between your team and every AI provider in the world. Every prompt is scanned in under 10ms, classified across 16 risk categories, and either forwarded, blocked, or quarantined — before it ever reaches the model.

BEFORE:  Your Team → OpenAI / Claude / Gemini / Llama
AFTER:   Your Team → Hound Shield Gateway → OpenAI / Claude / Gemini / Llama
                          ↓ 16-engine scan ↓
                   Blocked · Quarantined · Logged · Anchored

One deployment. 800+ models. Three compliance frameworks simultaneously: SOC 2, HIPAA, and CMMC Level 2.


Demo

Live demo: houndshield.com — interact with Brain AI (bottom right) for instant compliance answers.

Gateway in Action

// Before — raw OpenAI call, no compliance
const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });

// After — every prompt scanned, logged, compliance-enforced
const client = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
  baseURL: "https://houndshield.com/api/gateway/intercept",
});

// Same API. 800+ models. Three compliance frameworks. <10ms overhead.
const response = await client.chat.completions.create({
  model: "anthropic/claude-sonnet-4-6",
  messages: [{ role: "user", content: "Review this defense contract..." }],
});
// If the message contained FOUO/CUI data → blocked before reaching Claude

Performance Benchmarks

Detection Latency (p50 / p95 / p99)

Scan Type p50 p95 p99
Short prompt (<256 chars) 0.4ms 1.2ms 2.1ms
Medium prompt (<1K chars) 1.8ms 4.3ms 7.6ms
Long prompt (<4K chars) 3.9ms 7.8ms 9.4ms
Stream output (per chunk) 0.1ms 0.3ms 0.6ms

Measured on Vercel Edge Runtime. LRU cache hit rate >50% in production reduces these by 90%.

Detection Accuracy (internal evaluation, 10K prompt test set)

Category Precision Recall F1
CUI / FOUO 99.1% 97.8% 98.4%
HIPAA PHI (18 identifiers) 98.7% 99.2% 98.9%
API Keys & Credentials 99.6% 98.1% 98.8%
PII (SSN, passport, DL) 97.4% 96.9% 97.1%
Source Code / IP 94.2% 91.7% 92.9%

Why Local-Only Matters for CMMC

Most DLP tools scan your data in the cloud. For CMMC compliance, that's the problem — not the solution.

CMMC Level 2 control AC.L2-3.1.3 (CUI Flow Control) requires that Controlled Unclassified Information not flow to unauthorized destinations. Sending prompts containing CUI to a third-party scanning service — even a well-intentioned security vendor — is itself a CUI disclosure. Cloud DLP is architecturally incompatible with CMMC.

Hound Shield runs entirely within your infrastructure. The proxy sits on your network. Pattern matching runs in your process. Nothing leaves your perimeter except the sanitized prompt forwarded to the AI provider. The only external call Hound Shield makes is a SHA-256 hash of your license key alongside a prompt count — never the prompt content, never the matched patterns, never any data. Your SPRS score, your audit logs, your CUI — all stay inside your boundary.


Why Teams Choose Hound Shield

Challenge Old Approach Hound Shield
AI prompt data leaks Block all AI tools Scan every prompt in <10ms
Multi-framework compliance Three separate tools, three audits One gateway, all frameworks at once
Deployment friction Weeks of network rerouting, agent installs 15 minutes, one baseURL change
False positives Regex-only, blocks everything 16-engine hybrid AI+regex detection
Model access One model per approved tool 800+ models, all compliant
Audit evidence Manual screenshots and exports Cryptographic ledger, one-click PDF
SPRS scoring Consultant spreadsheet, $15K/assessment Live dashboard, 110 controls tracked
Security alerts Email, maybe Slack, Teams, Splunk, Sentinel — all at once

Key Features

800+ Models. Zero Compromise.

Access every major AI model — GPT-4o, Claude 4.6, Gemini 2.5 Pro, Llama 3.3, Mistral Large, Qwen3, Command R+, and 800 more — all through a single OpenAI-compatible endpoint. Every model, every prompt, fully scanned. Your team gets access to the best AI on the market. Your security team gets a clean audit trail for all of it.

// Switch any model without changing your code
const client = new OpenAI({
  baseURL: "https://houndshield.com/api/gateway/intercept",
  apiKey: process.env.OPENAI_API_KEY,
});

// GPT-4o, Claude 4.6, Gemini — same URL, all compliant
const response = await client.chat.completions.create({
  model: "anthropic/claude-sonnet-4-6",  // or gpt-4o, google/gemini-flash-1.5...
  messages: [{ role: "user", content: "..." }],
});

16-Engine Detection Matrix

Parallel scanning across 16 risk categories in a single sub-10ms pass:

Category What We Catch
Controlled Unclassified Info (CUI) FOUO markings, CUI//SP-PRVCY headers, DoD identifiers
Protected Health Information Patient names, diagnoses, MRNs, SSNs, insurance IDs
API Keys & Credentials sk-, AKIA, JWT tokens, OAuth secrets, DB connection strings
CAGE Codes & Contract Numbers Defense procurement, DD Form 254 references
Security Clearance Levels SECRET, TOP SECRET, SCI, SAP references
Export Controlled Data ITAR/EAR classifications, ECCN codes
Personally Identifiable Info Full names, addresses, DOBs, passport numbers
Financial Instruments Card numbers, IBAN, routing, SWIFT, wire instructions
Intellectual Property Patent filings, algorithms, proprietary formulas
Source Code & Schemas Proprietary logic, database schemas, system architecture
M&A Documentation Deal terms, target companies, valuations
Legal Privilege Attorney-client communications, litigation strategy
Biometric Data Facial recognition refs, fingerprint data, DNA
Geolocation Intelligence Military coordinates, classified facility locations
Network Infrastructure Internal IPs, firewall rules, VPN configs
Authentication Tokens OAuth flows, session tokens, API keys

AI Agent Framework Support

Hound Shield intercepts compliance across modern agentic architectures:

Framework Integration Method
Claude Code / Cursor Set gateway as base URL in .claude/settings.json
LangChain / LangGraph openai_api_base parameter
Goose (Block) OPENAI_BASE_URL environment variable
AgentScope Model provider config api_base
AutoGen base_url in OAI_CONFIG_LIST
LlamaIndex OpenAI(api_base=...) constructor

Every agent tool call, every LLM request — fully compliant.

Anthropic Managed Agents Integration

Hound Shield routes compliance scanning for Anthropic Managed Agents — composable cloud-hosted agent APIs. When Claude agents autonomously fetch data, invoke tools, or generate outputs, every LLM call passes through the Hound Shield 16-engine scan before reaching the model. No code changes required beyond the base URL.

Model Compliance Leaderboard

See exactly which AI models your team uses most — and how clean each one is. Hound Shield tracks compliance scores per model: how often each model triggers violations, what categories fire most, and which models hallucinate sensitive data. Choose the right model for regulated workloads with full evidence.

Real-Time Stream Scanning

Hound Shield scans the output stream token by token, not just the input. If an LLM starts generating a credit card number mid-response, the stream is truncated immediately before the full number is delivered. Real-time alerts fire before the data leaves the gateway.

Blockchain-Anchored Audit Trail

Every compliance event is:

  1. SHA-256 hashed and appended to a cryptographic chain (tamper-evident ledger)
  2. Anchored to Base L2 blockchain — permanent, immutable, publicly verifiable
  3. Exported instantly as CSV or JSON for auditors — filtered by date, risk level, or action

The cryptographic chain is dual-redundant: SHA-256 seed anchors + on-chain transactions. If any historical record is modified, the chain breaks. Instant tamper detection.

AES-256 Quarantine Vault

High-risk prompts are never stored in plaintext. They're:

  • Encrypted with AES-256 at rest
  • Isolated in a dedicated vault
  • Routed to a human reviewer via HITL approval workflow
  • Released or permanently deleted — never silently forwarded

Live SPRS Score Dashboard (CMMC)

Defense contractors see their NIST 800-171 Rev 2 score across all 110 controls in real time. Every AI intercept either helps or hurts your score. Watch it improve as you close gaps. Export as a C3PAO-ready PDF when it's time for your assessment. DoD scoring methodology v1.2.1.

Enterprise Alerting — Slack, Teams, Splunk & Sentinel

Compliance events reach your security team wherever they work:

Channel Integration What You Get
Slack Block Kit cards with severity colors Rich alerts, action buttons, multi-channel routing
Microsoft Teams Adaptive Cards v1.5 FactSet tables, severity icons, review links
Splunk HTTP Event Collector (HEC) NDJSON batch delivery, CIM-compatible fields
Azure Sentinel Log Analytics Workspace API HMAC-signed, KQL-queryable Hound ShieldCompliance_CL
CEF Common Event Format QRadar, ArcSight, syslog-compatible

Every critical alert routes to your SOC simultaneously. Zero additional configuration per-tool.

Sub-10ms Latency

Async generator streaming pipeline with 500-character scan windows, 256-character overlap, and LRU classification cache. Pattern matching short-circuits on the first CRITICAL detection. Your team never notices the firewall is there.


Quick Start

Option 1: Cloud (Recommended — 15 minutes)

  1. Sign up at houndshield.com
  2. Copy your gateway URL from the dashboard
  3. Change one line in your AI SDK:
// Before
const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });

// After — every prompt now scanned, logged, and compliance-enforced
const client = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
  baseURL: "https://houndshield.com/api/gateway/intercept",
});

That's it. No agents. No network changes. No training required.

Option 2: Self-Host (On-Prem / Air-Gap)

git clone https://github.com/thecelestialmismatch/Hound Shield.git
cd Hound Shield/compliance-firewall-agent

# Docker (recommended for production)
cp .env.example .env
# Fill required values — see Environment Variables table below
docker compose up -d

# Or: local development
npm install && npm run dev

Dashboard and gateway start at http://localhost:3000.


Integration Examples

Works with every OpenAI-compatible client. One URL. Every model.

Python

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["OPENAI_API_KEY"],
    base_url="https://houndshield.com/api/gateway/intercept",
)

# GPT-4o, Claude 4.6, Gemini — all compliant through one endpoint
response = client.chat.completions.create(
    model="anthropic/claude-sonnet-4-6",
    messages=[{"role": "user", "content": "Help me review this contract..."}]
)

LangChain

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="gpt-4o",
    openai_api_base="https://houndshield.com/api/gateway/intercept",
    openai_api_key=os.environ["OPENAI_API_KEY"],
)

Node.js / TypeScript

import OpenAI from "openai";

const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
  baseURL: "https://houndshield.com/api/gateway/intercept",
});

Goose (Block AI Agent Framework)

# ~/.config/goose/config.yaml
GOOSE_PROVIDER: openai
OPENAI_BASE_URL: https://houndshield.com/api/gateway/intercept
OPENAI_API_KEY: your-key-here

cURL

curl https://houndshield.com/api/gateway/intercept/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "x-api-key: $HOUNDSHIELD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "messages": [{"role": "user", "content": "Summarize this document..."}]
  }'

Environment Variables

Required

Variable Where to get it Purpose
NEXT_PUBLIC_SUPABASE_URL Supabase Settings > API Database URL
NEXT_PUBLIC_SUPABASE_ANON_KEY Supabase Settings > API Client auth
SUPABASE_SERVICE_ROLE_KEY Supabase Settings > API Server-side operations
OPENROUTER_API_KEY OpenRouter 800+ model access
NEXT_PUBLIC_APP_URL Your deployment URL Review links in alerts

AI & Detection

Variable Default Purpose
GEMINI_API_KEY Context-aware ML scanning (elevates regex findings)
BRAIN_AI_MODEL google/gemini-flash-1.5 Agent intelligence model

Notifications — Slack

Variable Purpose
SLACK_WEBHOOK_URL Default channel for all compliance alerts
SLACK_ALERT_WEBHOOK_URL Separate channel for HIGH/CRITICAL events only

Notifications — Microsoft Teams

Variable Purpose
TEAMS_WEBHOOK_URL Default Teams channel
TEAMS_ALERT_WEBHOOK_URL Separate channel for HIGH/CRITICAL events

SIEM — Splunk

Variable Default Purpose
SPLUNK_HEC_URL HEC endpoint (https://splunk.company.com:8088/services/collector/event)
SPLUNK_HEC_TOKEN HEC authentication token
SPLUNK_INDEX houndshield Target Splunk index
SPLUNK_SOURCETYPE _json Set to cef for CEF format
SPLUNK_HOST houndshield-gateway Host field value

SIEM — Azure Sentinel

Variable Purpose
SENTINEL_WORKSPACE_ID Log Analytics Workspace ID
SENTINEL_SHARED_KEY Primary or secondary workspace key
SENTINEL_LOG_TYPE Custom log type (default: Hound ShieldCompliance)

Blockchain Anchoring

Variable Default Purpose
BLOCKCHAIN_ANCHOR_PRIVATE_KEY 0x-prefixed Base L2 wallet key
BASE_RPC_URL https://mainnet.base.org Production RPC
BASE_SEPOLIA_RPC_URL https://sepolia.base.org Testnet RPC

Billing & Email

Variable Purpose
STRIPE_SECRET_KEY Stripe secret key
STRIPE_WEBHOOK_SECRET Stripe webhook signing secret
RESEND_API_KEY Transactional email

Security

Variable Purpose
ENCRYPTION_KEY 64-char hex key for AES-256 quarantine vault

Architecture

┌─────────────────────────────────────────────────────────────┐
│                     Your AI Clients                          │
│   Python · Node.js · LangChain · Copilot · Goose · Any SDK  │
└─────────────────────┬───────────────────────────────────────┘
                      │ HTTPS — OpenAI-compatible API
┌─────────────────────▼───────────────────────────────────────┐
│              Hound Shield Gateway  (Next.js 15 Edge)               │
│                                                              │
│  ┌───────────────────────────────────────────────────────┐  │
│  │         16-Engine Detection Matrix  (<10ms)            │  │
│  │  CUI · PHI · PII · Credentials · ITAR · IP · Fin...  │  │
│  │  Regex (CRITICAL-first) + Gemini Flash ML scan         │  │
│  │  LRU result cache · short-circuit on CRITICAL          │  │
│  └──────────────┬──────────────────────┬──────────────────┘  │
│                 │ PASS                 │ BLOCK / QUARANTINE  │
│  ┌──────────────▼─────────────┐       │                     │
│  │  Stream Proxy + Scanner    │       │                     │
│  │  Real-time token scanning  │       │                     │
│  │  Truncate on CRITICAL hit  │       │                     │
│  └──────────────┬─────────────┘       │                     │
└─────────────────┼─────────────────────┼─────────────────────┘
                  │                     │
  ┌───────────────▼──────┐  ┌───────────▼──────────────────┐
  │  800+ Model Providers │  │   AES-256 Quarantine Vault   │
  │  (via OpenRouter)     │  │   + HITL Review Workflow     │
  │  OpenAI · Anthropic   │  └──────────────────────────────┘
  │  Google · Mistral...  │
  └───────────────┬───────┘
                  │
  ┌───────────────▼──────────────────────────────────────────┐
  │                   Audit & Alert Layer                      │
  │  SHA-256 Seed Chain → Base L2 Blockchain Anchor           │
  │  Slack · Teams · Splunk HEC · Azure Sentinel · CEF        │
  └───────────────┬──────────────────────────────────────────┘
                  │
  ┌───────────────▼──────────────────────────────────────────┐
  │              Hound Shield Dashboard  (Next.js 15)               │
  │  SPRS Score · Compliance Events · Model Leaderboard       │
  │  PDF Reports · API Keys · Team Management · Audit Export  │
  └──────────────────────────────────────────────────────────┘

Multi-Agent System

Five specialized compliance agents powered by the Hound Shield ReAct orchestrator:

┌─────────────────────────────────────────────────────────────────┐
│                  Hound Shield Multi-Agent System                       │
│                                                                  │
│  AgentRouter — classifies task → routes to best agent           │
│  ┌─────────────┐ ┌─────────────┐ ┌─────────────┐               │
│  │  Scanner    │ │  Policy     │ │  Research   │               │
│  │  Agent      │ │  Agent      │ │  Agent      │               │
│  │  Real-time  │ │  Rule eval  │ │  Reg docs   │               │
│  │  PHI/CUI    │ │  Exception  │ │  NIST/CMMC  │               │
│  │  detection  │ │  requests   │ │  HIPAA/SOC2 │               │
│  └─────────────┘ └─────────────┘ └─────────────┘               │
│  ┌──────────────────────┐ ┌──────────────────┐                  │
│  │  Optimization Agent  │ │  Auditor Agent   │                  │
│  │  Latency profiling   │ │  SPRS scoring    │                  │
│  │  Cost/model routing  │ │  SHA-256 chain   │                  │
│  │  P50/P95/P99 metrics │ │  PDF evidence    │                  │
│  └──────────────────────┘ └──────────────────┘                  │
│                                                                  │
│  Shared: agentMemory (conversation + insights + preferences)     │
│  Brain AI knowledge index (queryKnowledge → context injection)   │
└─────────────────────────────────────────────────────────────────┘
Agent Model Tools Use
Scanner Gemini 2.0 Flash compliance-scan, data-query, knowledge-base PHI/CUI/PII detection
Policy Claude 3 Haiku compliance-scan, data-query, web-search Rule eval + exceptions
Research Llama 3.3 70B web-search, web-browse, knowledge-base Regulation research
Optimizer Gemini 2.0 Flash data-query, generate-chart, code-execute Latency + cost tuning
Auditor DeepSeek V3 data-query, file-analyze, generate-chart SPRS scoring + reports

Brain AI — Knowledge Ingestion Pipeline

IngestionSource (URL or content)
    │
    ▼ fetchContent (10s timeout, 500KB limit)
    │
    ▼ normalizeContent (strip HTML/markdown noise)
    │
    ▼ chunkText (800-char chunks, 100-char overlap)
    │
    ▼ contentHash (SHA-256 dedup — never re-index)
    │
    ▼ extractTags (compliance · security · agents · performance)
    │
    ▼ knowledgeStore (global Map — zero external deps)
    │
    ▼ queryKnowledge (keyword scoring → agent context injection)

API: POST /api/brain-ai/ingest · GET /api/brain-ai/ingest?q=<query>&domain=compliance

Performance Enforcement

Every gateway operation runs inside withBudget() — hard latency gates with P50/P95/P99 tracking:

Operation Budget Action on Breach
Regex scan 5ms WARN log + counter increment
ML/semantic scan 10ms WARN log + counter increment
Gateway total 50ms ERROR log + violation event
First token 200ms WARN (provider-side)

SHA-256 audit chain: every log entry hashes {id, timestamp, event_type, actor, data, prevHash} — any tampering breaks all subsequent hashes.

API: GET /api/gateway/metrics — live P50/P95/P99 + exceedance rates + health status

Stack:

Layer Technology
Frontend Next.js 15, React 19, TypeScript 5.8, Tailwind CSS, Framer Motion
Backend Next.js API routes (Node.js runtime), Supabase PostgreSQL + RLS
AI Engine OpenRouter (800+ models), Gemini Flash (ML scanning)
Auth Supabase Auth — Google, GitHub, Microsoft OAuth
Billing Stripe — 5 subscription tiers
Alerts Slack Block Kit, Teams Adaptive Cards, Splunk HEC, Azure Sentinel
Blockchain Viem + Base L2 (Ethereum L2) — ~$0.001 per anchor
Email Resend — transactional notifications
Deployment Vercel (cloud) · Docker + nginx (self-hosted)
Multi-Agent 5 specialized ReAct agents via OpenRouter (free + paid models)
Brain AI SHA-256 dedup knowledge index, keyword search, domain tagging

Pricing

Tier Price AI Scans/mo Users Highlights
Starter Free 1,000 1 Dashboard · 16-engine detection
Pro $199/mo 50,000 10 SOC 2 + HIPAA · Slack + Teams
Growth $499/mo 250,000 25 + CMMC L2 · SIEM connectors · PDF reports
Enterprise $999/mo Unlimited Unlimited + Blockchain anchoring · White-label · On-prem
Agency / MSP $2,499/mo Unlimited Multi-tenant Full platform resale · Custom branding

All paid plans include a 14-day free trial. No credit card on Starter.


Compliance Coverage

SOC 2 Type II

Continuous monitoring of all AI traffic with cryptographic, tamper-evident audit trail. Instant evidence export (CSV + JSON). Controls mapped to AICPA Trust Service Criteria. Blockchain-anchored records available on Enterprise.

HIPAA

Automatic PHI detection before any AI provider ever sees the request. HIPAA-compliant logging with SHA-256 hashing (no PHI stored in plaintext). AES-256 quarantine for high-risk prompts. Business Associate Agreement available on Growth+.

CMMC Level 2 — NIST 800-171 Rev 2

All 110 controls tracked in real time. Live SPRS score using DoD methodology v1.2.1. CUI detection and blocking (FOUO, ITAR, CAGE codes, clearance levels). Automated evidence collection. C3PAO assessment coordination on Enterprise.


Industry Verticals

Vertical What Hound Shield Protects
Defense Contractors CUI, CAGE codes, CMMC enforcement, clearance level references
Healthcare Systems PHI, HIPAA audit trail, EHR prompt sanitization
Technology Companies SOC 2 evidence, credential leaks, IP protection
Legal & Finance Attorney-client privilege, PCI-DSS, M&A confidentiality
Government Agencies ITAR/EAR, classified data, FISMA alignment
Five Eyes Partners Multi-jurisdiction compliance, allied-nation data residency

Roadmap

Shipped

  • AI gateway proxy — OpenAI-compatible, 800+ models via OpenRouter
  • 16-engine CUI/PHI/PII detection matrix
  • Real-time stream scanning — token-level truncation on CRITICAL
  • LRU classification cache — >50% latency reduction on repeated content
  • SPRS score calculator — all 110 NIST 800-171 controls
  • Cryptographic audit ledger — SHA-256 seed chain
  • AES-256 quarantine vault + HITL review workflow
  • PDF compliance reports — SOC 2, HIPAA, CMMC
  • Stripe billing — 5-tier subscription management
  • Multi-framework dashboard — SOC 2, HIPAA, CMMC simultaneously
  • Slack native integration — Block Kit, multi-channel routing, retry
  • Microsoft Teams integration — Adaptive Cards v1.5
  • SIEM connector — Splunk HEC (NDJSON batch + CEF)
  • SIEM connector — Azure Sentinel (HMAC-signed, KQL-queryable)
  • Blockchain-anchored audit trail — Base L2, ~$0.001/anchor
  • On-prem Docker deployment — multi-stage build, non-root, health checks
  • Model compliance leaderboard — per-model violation rate tracking
  • Async generator pipeline — parallel pattern matching, short-circuit
  • Brain AI — local compliance FAQ engine + LLM fallback
  • Goose / AgentScope multi-agent framework support
  • Anthropic Managed Agents gateway integration

In Progress

  • Browser extension — Chrome, Edge (intercept browser-native AI tools)
  • Mobile app — iOS + Android (Hound Shield dashboard + push alerts on the go)
  • MCP Server endpoint — model context protocol for Claude Desktop, Cursor

Planned

  • Ollama / local model support — full air-gap mode
  • SIEM connector — Elastic/ELK Stack
  • SAML / SSO — Okta, Azure AD, JumpCloud
  • Webhook delivery receipts and dead-letter queue
  • Custom pattern library — org-specific regex + ML rules
  • Zero-trust mode — deny by default, allowlist per team
  • REST API v1 — programmatic access for MSP/consultant integrations

SIEM Query Examples

Azure Sentinel (KQL)

// All blocked prompts in the last 24 hours
Hound ShieldCompliance_CL
| where action_s == "BLOCKED"
| where TimeGenerated > ago(24h)
| project TimeGenerated, user_id_s, risk_level_s, entity_types_s, provider_s
| order by TimeGenerated desc

// CRITICAL events by user
Hound ShieldCompliance_CL
| where risk_level_s == "CRITICAL"
| summarize count() by user_id_s, bin(TimeGenerated, 1h)
| render timechart

Splunk SPL

index=houndshield sourcetype=_json action=BLOCKED
| stats count by user_id, risk_level, entity_types
| sort - count

index=houndshield risk_level=CRITICAL
| timechart span=1h count by provider

Security

We treat security issues seriously. Do not open a public GitHub issue for vulnerabilities.

Contact: info@houndshield.com

See SECURITY.md for our responsible disclosure policy.

Security Architecture

  • All secrets accessed via environment variables — never hardcoded
  • Prompt text never stored in plaintext — SHA-256 hashed only
  • AES-256 encryption for quarantined content
  • Non-root Docker user (nextjs:nodejs)
  • Rate limiting at middleware layer (60 req/min default)
  • HMAC-SHA256 for all outbound SIEM requests
  • Base L2 blockchain for tamper-proof audit evidence

Acknowledgements

Hound Shield integrates with and is informed by the following open-source ecosystems:


Contributing

  1. Fork the repo and create a branch: git checkout -b feat/your-feature
  2. Make changes. Run cd compliance-firewall-agent && npm run build — must pass.
  3. Tests are required: npm test -- --coverage must show 80%+ coverage on changed files.
  4. Open a PR against main. The code-reviewer agent runs on every PR.

Do not open PRs that:

  • Reduce CUI/PHI detection pattern count below 16
  • Modify SPRS scoring deductions without citing the NIST 800-171 Rev 2 control
  • Add external network calls that could transmit prompt content

Security issues: email info@houndshield.com — do not open public issues for vulnerabilities.


License

MIT — free to self-host, fork, and modify. Commercial use of the hosted service at houndshield.com requires a paid plan.


Built for the 80,000+ defense contractors facing CMMC enforcement in 2026.

And every healthcare system, law firm, and tech company that's realized their team is already using AI — whether they approved it or not.


Website Docs Pricing Email


Engineering uncompromising security for the next generation of artificial intelligence.

About

Local-only AI compliance firewall. Stop CUI & PHI leaks to ChatGPT, Copilot & Claude in <10ms. CMMC Level 2, HIPAA, SOC 2 enforced simultaneously.

Topics

Resources

License

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors