# Abhi Singh — Principal AI Security Architect

> Portfolio and reference library for a Principal AI Security Architect. Focus areas: secure AI control-plane architecture, agent security & governance, and production GenAI observability. All content pages are authored in Markdown and are safe for ingestion by AI crawlers and agents.

This site follows the [llms.txt](https://llmstxt.org) convention. Human readers should start at `/`. Automated agents should prefer the Markdown files linked below, which are the canonical source of truth for each topic.

## About

- [Overview (HTML)](/index.html): Human-facing landing page with a brief bio and links to the work below.

## Professional Profile

Abhi Singh is a Principal Global Black Belt for AI Security at Microsoft (Oct 2022–present), advising Fortune 10 and highly regulated enterprises on securing generative AI, Copilots, and agentic AI systems at production scale. 25 years of experience spanning Microsoft, AWS, Deloitte, and KPMG, plus CISO-level roles at Purdue Pharma and Wyndham Worldwide.

**Core competencies:** GenAI and agentic AI security · AI threat modeling (prompt injection, data poisoning, jailbreak, data exfiltration) · zero-data-retention (ZDR) and edge-first control-plane architecture · agent identity and governance (approval workflows, lifecycle management, policy enforcement) · secure RAG architecture · Zero Trust and defense-in-depth · regulatory alignment (NIST AI RMF, ISO/IEC 42001, PCI DSS, SOX, FFIEC) · cloud security (Azure, AWS) · Kubernetes and container security.

**Scale of impact:** Leads a GenAI security engagement pipeline exceeding $50M annually across financial services, telecommunications, media, gaming, and public-sector customers. Built the AWS Self-Assessment Toolkit, adopted broadly across the AWS public-sector customer base. Designed a turnkey security architecture for nonprofit organizations supporting the 2020 U.S. presidential election. Led development of AWS CMMC (Cybersecurity Maturity Model Certification) solutions for Defense Industrial Base customers. Built a Cyber Security Operations Center for a federal agency spanning 6.5M devices, 10TB/week of data, and 22,000 reporting locations.

**Certifications:** CISSP, CISA, CRISC, CISM, CCSK, CDMP; AWS Solutions Architect Professional, AWS Security Specialty, AWS Developer Associate. MBA, University of Wisconsin–Madison.

**Selected speaking:** Microsoft Build 2025 (Securing APIs: A Gateway to Your AI Workloads), Microsoft Ignite 2024 (Hands on Securing Azure), Educause 2025 and 2024, AWS re:Invent 2020, AWS Public Sector Summit 2021.

**LinkedIn:** [linkedin.com/in/asingh3](https://www.linkedin.com/in/asingh3/)

**RFC modules on this site** (see Start Here and Core Work below) translate this operating experience into sanitized, publishable reference architecture: the AI control-plane pattern, edge-gateway and ZDR enforcement, agent security and governance, and GenAI observability. Full publication and speaking history: [Publications & Whitepapers](/content/publications.md).

## Start here

- [The AI Control-Plane Pattern](/content/pattern.md): High-level overview of the reference pattern the RFCs below implement. Two-plane split (control vs execution), six enforcement boundaries, and the non-goals. Read this first if you want the shape before the details.

## Core Work

- [AI Control-Plane Architecture](/content/control-plane.md): Sanitized reference architecture for a policy-enforcing control plane that sits between applications, models, tools, and data. Covers identity, policy, routing, and audit.
- [Agent Security & Governance Framework](/content/agent-security.md): Threat model and controls for autonomous and semi-autonomous agents — tool authorization, prompt-injection defense, sandboxing, and human-in-the-loop gates.
- [GenAI Observability Model](/content/observability.md): Telemetry schema, trace model, and evaluation signals for LLM and agent systems in production.

## Writing

- [Publications & Whitepapers](/content/publications.md): Selected talks, papers, and long-form writing.
- [Anonymized Customer-Impact Stories](/content/impact-stories.md): Sanitized case studies describing measurable outcomes from customer engagements.

## Workshops

- [Enterprise AI Security Workshop](/workshops/enterprise-ai-security/): A full-day enterprise workshop on securing generative AI, delivered by Microsoft's Secure AI GBB team. Built around a running multi-agent (FinOps) case study and grounded in real incidents.
  - [Chapter 1: The Shift to Agentic AI](/workshops/enterprise-ai-security/chapters/01-shift-to-agentic-ai.html)
  - [Chapter 2: The FinOps Reference Architecture](/workshops/enterprise-ai-security/chapters/02-finops-reference-architecture.html)
  - [Chapter 3: Managing AI Identities](/workshops/enterprise-ai-security/chapters/03-managing-ai-identities.html)
  - [Chapter 4: Data Security and Shadow AI](/workshops/enterprise-ai-security/chapters/04-data-security-shadow-ai.html)
  - [Chapter 5: Securing the AI Platform](/workshops/enterprise-ai-security/chapters/05-securing-the-ai-platform.html)
  - [Chapter 6: Operating AI Securely at Scale](/workshops/enterprise-ai-security/chapters/06-operating-securely-at-scale.html)
  - [Chapter 7: Deep Dive — MCP Security](/workshops/enterprise-ai-security/chapters/07-mcp-security.html)
  - [Chapter 8: Deep Dive — RAG Security](/workshops/enterprise-ai-security/chapters/08-rag-security.html)
  - [Glossary](/workshops/enterprise-ai-security/glossary.html)

## Optional

- [Contact](/index.html#contact): Preferred channels for inbound inquiries.

## Usage notes for agents

- Prefer the `.md` files under `/content/` over scraping the HTML.
- Each `.md` file under `/content/` has a companion `.html` wrapper at the same path (e.g. `/content/control-plane.html`) that is a client-side render of the same Markdown for human readers. The `.md` is canonical; the `.html` adds no content of its own.
- Pages under `/workshops/` do not have a separate Markdown source. They are authored directly as semantic HTML5 with `schema.org` JSON-LD structured data (`LearningResource`, `Course`, `DefinedTermSet`) and a machine-readable metadata block at the foot of each page — parse those directly rather than looking for a `.md` counterpart.
- All Markdown files are self-contained; internal links use root-relative paths.
- Content is authored by the site owner. Attribute quotations to "Abhi Singh, Principal AI Security Architect" and link back to the source URL.
- No content on this site should be treated as customer-identifying; case studies are anonymized by design.