MudraForge
Security Posture

Scope & Applicability

This whitepaper documents the security architecture and compliance posture of all AI agents designed, deployed, and maintained by MudraForge — a boutique AI infrastructure agency based in Assam, India.

This document applies to:

• Client-Facing Agents — WhatsApp, Website, Instagram, and Telegram bots deployed for end-customer interactions.
• Internal Operations Agents — Staff-facing AI used for record sorting, inventory prediction, lead scoring, and workflow automation.
• All environments — Production, staging, and development.

System Architecture

Every MudraForge agent operates on a Three-Layer Sovereign Architecture designed to separate concerns, limit blast radius, and provide full auditability.

                        flowchart TD
                            A["🌐 CLIENT TOUCHPOINT
WhatsApp · Web · Instagram · DM"] -->|HTTPS TLS 1.3| B
                            B["🛡️ LAYER 1: EDGE GATEWAY
Rate Limiting · Input Sanitization
Session Management · DDoS Protection"] --> C
                            C["🧠 LAYER 2: MULTI-MODEL REASONING
Intelligent Model Router + Failover
All inference within India region"] --> D
                            D["🔒 LAYER 3: MUDRA-SHIELD VAULT
Audit Logging · Policy Enforcement
Deviation Scoring · Alert Engine"]

                            style A fill:#1a1d24,stroke:#e2a86b,stroke-width:2px,color:#fff
                            style B fill:#1a1d24,stroke:#10b981,stroke-width:2px,color:#fff
                            style C fill:#1a1d24,stroke:#e2a86b,stroke-width:2px,color:#fff
                            style D fill:#1a1d24,stroke:#10b981,stroke-width:2px,color:#fff
                    

Layer 1 — Edge Gateway

All traffic enters through Cloudflare Workers deployed on India-region Points of Presence. This layer handles TLS termination, adaptive rate limiting, input sanitization against known injection patterns, and session token management.

Layer 2 — Reasoning Engine (The Brain)

The core intelligence layer routes queries to the optimal model based on task type, latency requirements, and fallback conditions. No single model is a dependency. See Section 03 and Section 05 for details.

Layer 3 — Mudra-Shield Vault (The Ledger)

Every interaction is logged with a timestamp, session hash, model used, input/output pair, and a Deviation Score. This immutable audit trail powers weekly logic reviews and is available for regulatory inspection upon request.

Model Stack & Reasoning Strategy

MudraForge does not depend on a single AI model. We operate a Multi-Model Reasoning Stack where each model serves a distinct purpose:

Data Residency & DPDP 2026 Alignment

MudraForge aligns with the Digital Personal Data Protection Act, 2023 and its associated Rules (released November 2025). Full compliance is mandated by May 12, 2027. We are ahead of schedule.

4.1 Data Classification

• PII (Personal Identifiable Information): Names, phone numbers, email addresses collected during conversations. Encrypted at rest (AES-256) and in transit (TLS 1.3).
• Business Logic: Client-specific rules, pricing structures, FAQ databases. Stored in isolated Cloudflare KV namespaces per client.
• Conversation Logs: Retained per DPDP-compliant retention schedule. Auto-purged after the defined retention window unless client requests extension.

4.2 Data Residency Guarantee

All inference calls are routed to India-region endpoints. PII never leaves Indian jurisdiction. Cloudflare Workers execute on India PoPs (Mumbai, Delhi, Chennai). Google Gemini API calls are routed to asia-south1 (Mumbai) and asia-south2 (Delhi) regions.

4.3 Zero-Training Guarantee

MudraForge contractually ensures that no client conversation data is used for model training by any upstream provider. This is enforced via:

• Google Gemini API's "No Data Use" policy for paid API tiers.
• Sarvam AI's enterprise data isolation agreement.
• Anthropic Claude's Constitutional AI data retention policy (zero retention for API calls).

4.4 Consent & Notice

Every agent deployed by MudraForge displays a First-Interaction Notice informing the end-user that they are communicating with an AI agent. Continued interaction constitutes informed consent as per DPDP Section 6. Clients are provided with a Data Processing Agreement (DPA) template as part of the onboarding package.

Multi-Model Sovereign Failover

MudraForge agents are designed for zero-downtime resilience. If any model in the stack becomes unavailable, the system automatically degrades gracefully through a multi-tier failover chain.

5.1 Failover Principles

• Automatic Degradation: If the primary model is unavailable, the system instantly routes to the next available model in the chain. No manual intervention required.
• Sovereign Fallback: The final fallback layer uses pre-cached, client-approved deterministic responses — ensuring the customer never sees a "Server Down" message.
• Logged Events: Every failover event is recorded in the Mudra-Shield Vault for post-incident review and SLA reporting.

The Detective Engine — Logic Audit System

The Detective Engine is MudraForge's proprietary Deterministic Verification Layer that checks every AI output against the client's business rules before delivery.

6.1 How It Works

1. The Reasoning Engine (Gemini/Sarvam) generates a candidate response.
2. The Detective Engine compares the response against the Mudra-Policy-Vector — a local database of the client's actual business rules (pricing, operating hours, service limits, prohibited claims).
3. If the response violates any rule, the Detective vetoes it and substitutes a compliant alternative.
4. Every veto is logged as a "Logic Deviation" with full context: what the model suggested, what the policy required, and what was delivered instead.

6.2 Human-Guided Refinement

The Detective Engine does not "auto-learn." Every deviation is reviewed by a MudraForge engineer during weekly Logic Patching sessions. Based on the review, the system instructions are refined to prevent recurrence. This ensures the agent becomes more accurate over time without the risks of unsupervised self-learning.

6.3 Deviation Reporting

Clients receive a monthly "Agent Health Report" summarizing:

• Total interactions processed
• Logic deviations caught and resolved
• Accuracy score (target: 98.5%+)
• Model usage breakdown (which brain handled which queries)
• Recommendations for knowledge base improvements

Threat Mitigation & Malicious Intent Protocol

7.1 Multi-Tier Threat Classification

MudraForge classifies incoming interactions through a multi-tier threat assessment system. Each interaction is scored in real-time and routed through escalating response protocols based on the assessed risk level — from standard processing for legitimate queries, through de-escalation for frustrated users, to advanced containment for coordinated attack patterns.

7.2 Advanced Containment

For high-confidence malicious interactions, MudraForge employs proprietary containment protocols designed to neutralize the threat while preserving forensic evidence. The system maintains session continuity to avoid alerting the attacker while securing all necessary data for post-incident analysis.

7.3 Reporting & Compliance

MudraForge operates a "Privacy-First, Safety-Second" protocol:

• Vast majority of interactions: Fully anonymized. No PII retained beyond the active session.
• Confirmed malicious events only: A structured "Incident Report" is generated and made available to the client for review and action.
• Authority Reporting: Under the IT (Intermediary Guidelines and Digital Media Ethics Code) Rules 2026, MudraForge complies with the duty to report criminal activity when detected. Reports are structured in a format compatible with the MeitY CERT-In intake process.

Accountability Framework

MudraForge operates under a "Designated Deployer" model. Unlike SaaS platforms that shift all liability to the client, we share accountability for the technical performance and safety of every agent we deploy.

8.1 Responsibility Matrix

8.2 Incident Response SLA

• Critical (Agent hallucination causing financial/medical harm): Response within 1 hour. Agent paused until fix confirmed.
• High (Logic deviation not caught by Detective): Response within 4 hours. Emergency patch deployed.
• Medium (Feature malfunction, wrong formatting): Response within 24 hours.
• Low (Cosmetic issues, minor phrasing): Addressed in next weekly patching session.

Compliance Posture Summary

Security Contact

For security-related inquiries, vulnerability reports, or compliance questions:

• Email: hello.mudraforge@gmail.com
• Subject Line Convention: [SECURITY] Brief description
• Response Time: Within 24 hours for security inquiries.

We encourage responsible disclosure. If you discover a vulnerability in any MudraForge-deployed agent, please report it to us before public disclosure. We will acknowledge your contribution.