Wayam Manthan · Security & Trust
Security, deployment & governed agent architecture
Enterprise agentic AI requires more than model safety prompts: it demands bounded VPC deployment, zero data retention, customer-controlled keys, and immutable audit logs.
- Model retention
- Zero days
- No training on customer data
- Deployment
- Customer boundary
- Azure, AWS or on-premises
- Transit and rest
- TLS 1.3, AES-256
- Customer-managed keys
- Act layer
- Human gated
- Dual-custody approval
Core architectural protections
Tenant isolation & deployment boundaries
Single-tenant container boundary running within your Azure/AWS tenant or on-premises infrastructure.
- No multi-tenant data bleed: agent state, document vectors, and tool execution environments run isolated per customer boundary.
- Support for Azure virtual networks (VNet), AWS PrivateLink, and air-gapped on-premises deployments using local models (vLLM, NVIDIA NIM).
- Egress controls strictly whitelist customer-approved tool endpoints and external system APIs.
Zero model data retention & privacy
Customer prompts, retrieved documents, and tool outputs are never used to train foundation models.
- Contractually bound commercial enterprise model endpoints with zero-retention (ZDR) guarantees.
- Customer data is discarded immediately after token completion; no caching or evaluation storage by LLM vendors.
- Strict data classification gates preventing confidential or regulated tokens from exiting approved geographical regions.
Encryption & customer-managed keys (BYOK)
End-to-end cryptographic protection in transit and at rest with customer-owned keys.
- TLS 1.3 enforced for all transit connections with forward secrecy and strict HSTS.
- AES-256 encryption for all data at rest, including vector embeddings and session metadata.
- Customer-managed keys (CMEK) via Azure Key Vault or AWS KMS; revoke access at any time.
Identity, RBAC & dual-custody approval
Fine-grained permissions and human approval gates before consequential actions execute.
- Enterprise SSO integration supporting SAML 2.0, OpenID Connect (OIDC), Azure AD / Entra ID, and Okta.
- Role-based access control (RBAC) separating Ingest/Reason visibility from Act-layer write authority.
- Configurable dual-custody approval workflows for irreversible ERP updates or financial transactions.
Immutable cryptographic audit logging
Complete traceability of every reasoning step, tool call, policy check, and human decision.
- Structured append-only trace log recording inputs, retrieved citations, prompt hashes, tool payloads, and approval signatures.
- Trace ID correlation linked directly to customer SIEM (Splunk, Microsoft Sentinel, Datadog).
- Tamper-evident hashing guaranteeing non-repudiation for audit and compliance teams.
Compliance & regulatory framework crosswalks
Aligned directly with international standards and national data protection regulations.
- India Digital Personal Data Protection (DPDP) Act 2025: purpose limitation, consent receipts, and erasure workflows.
- NIST AI Risk Management Framework (AI RMF 1.0) & GenAI Profile: GOVERN, MAP, MEASURE, MANAGE mapping.
- OWASP Top 10 for LLM applications: automated mitigations for prompt injection, sensitive data leakage, and excessive agency.
- ISO/IEC 42001 (Artificial Intelligence Management System): governance policies currently in formal audit review.
Governance sample pack
Enterprise procurement, compliance, and risk committees receive inspectable documentation before any pilot commences. Here is what is included in our standard governance artifact set:
Architecture specification, model provenance, acceptable use boundaries, and known failure modes.
Operational risk scoring, autonomy ceilings, and required control gates mapped to business impact.
50-run synthetic evaluation suite with accuracy thresholds, citation checks, and drift alerts.
Operating procedures for edge-case hand-off, operator override, and model policy update cadence.