Security Architecture

Defense in Depth

Security is not a layer — it's the foundation. Every component fails closed. Every action is auditable. Every model is governed.

⟡ AMGL Guard

Pre-inference policy enforcement. 23 rules in ActiveMirrorOS. Blocks prompt exfiltration, role injection, social engineering, jailbreaks.

23
Guard Rules
12,096
Lines in Control Plane
Fail-Closed
Default Behavior

🔒 MirrorGate

Application-layer safety proxy. Real-time safety scoring. Sits between user and model — every request is evaluated before inference begins.

⛨ Kavach — Sovereign AI Shield

Consumer-facing scam detection and digital safety for India. 16 shipped capabilities including deepfake detection, voice clone analysis, QR code scanning, and SMS auto-scanning.

Kavach Factory Manifest

Factory manifest for Kavach sovereign AI shield. 4 parallel Claude Code agents building core agent, Telegram bot, scam detection engine, and Bhashini+Govt services. All cwd /Users/mirror-admin/repos/kavach.

Kavach MVP — API + Telegram Bot

Kavach sovereign AI shield running. 3,266 lines: scam detection (15 categories, 52 tests), govt schemes (10), Bhashini (12 languages), Telegram bot, Ollama llama3.2:3b. Voice endpoint wired, Ollama health check on startup.

Kavach Web UI

Browser-based scam scanner with quick scan and deep scan, example messages, risk card display

Kavach Hindi Support

Hindi risk responses for HIGH/MEDIUM/LOW scam results via /lang hi command

Kavach LaunchAgents

LaunchAgents for auto-restart on reboot: server on :8790 and Telegram bot with lock-clearing starter

Kavach DPDP Compliance

DPDP Act compliance: disclaimers on all risk cards, consent tracking, 7-day data retention, privacy rights API (/export, /delete, /consent), privacy policy, ToS

Kavach WhatsApp Bot

WhatsApp Business Cloud API integration: webhook verification, text+image message handling, OCR reuse, rate limiting, Hindi+English templates

Kavach Verification Lattice

Crowdsourced threat DB + Google Safe Browsing: URL/phone/UPI reputation checks, anti-poisoning (min 3 reports), 233 govt whitelist domains, 70 seeded threats, Chakshu/cybercrime.gov.in reporting links

Kavach UPI Checker + Web UI v2

Indian UPI ID validation and fraud detection, phone number reputation, server.py upgraded to 18 endpoints, risk_engine.py wired to lattice with graceful degradation

Kavach v0.4 Endpoints

v0.4 additions: event log, PNG risk cards, Android call screen bridge

Chetana v3 Security Hardening

XSS esc(), rate limiter, analytics JSONL, feedback endpoint, Ollama fallback, OG card

Chetana Media Deepfake Analyzer

898-line EXIF + Ollama vision deepfake detection for images/videos, 30/70 scoring

Chetana Discovery + Trust Surface Hardening

Honest timing/privacy claims, official resources hub, Atom feed, AI-discovery metadata, and live-route hardening for chetana.activemirror.ai.

Chetana Legal Renderer + Truth Surface Repair

Fixes live legal markdown rendering, truthful landing-page browser-model copy, and internal legal links for Chetana's public surface

Chetana Public HEAD Support

Adds explicit HEAD support for Chetana public pages so crawlers, probes, and CDN checks get 200 alongside normal GET responses

Chetana UI v3 — 10 Tabs

Added QR Scan, Voice Check, Link Check tabs + JS handlers to Chetana UI (now 10 tabs)

🔐 Infrastructure Security

System-level security hardening, monitoring, and self-healing.

🔒

chetana.activemirror.ai DNS

DNS CNAME + cloudflared ingress for chetana.activemirror.ai → localhost:8790

🔒

System Snapshot (30min)

Full state capture every 30min: 71 repos git status, 7 ports, launchd agents, key file hashes. 48hr retention.

🔒

Repo Security Audit

Audit + privatize 13 security-sensitive repos (MirrorGate, mirrorbrain-api, MirrorBrain-Mobile, ActiveMirrorOS, etc.)

🔒

SERVICE_REGISTRY audit

Added 3 missing services: MirrorBalance :8400, Pulse Swarm Hub :7771, Ollama Proxy :11435. 42→45 services.

🎯 Red-Team Testing

175 attacks tested across 5 categories (December 2025). Prompt exfiltration, role injection, meta-instruction, social engineering, jailbreak patterns. Vulnerabilities found, patched, verified.