Show Notes
Should you block AI agents? This episode answers it from the identity chair, and it starts with MCP authorization explained from zero. If you've heard "MCP authentication" and "OAuth" in the same breath all year and never quite connected them, this episode does it step by step.
We starts with OAuth 2.1 from scratch (what the roles are, what a token actually is, why it's not login), then MCP from scratch (what the Model Context Protocol is, local versus remote servers, why an AI agent is unlike any API client you'vesecured before), then walks the full MCP authorization handshake with the July 28, 2026 spec open: the 401, Protected Resource Metadata, Client ID Metadata Documents, PKCE, resource indicators, issuer validation, audience binding,step-up scopes, and the two rules that matter most. Then why this lands on the identity leader's desk, and the question the CISO will actually ask: do we let agents in at all? We walk the block-or-admit debate, the August 2026 Ninth Circuit ruling in Amazon v. Perplexity, why the fraud bouncer can't tell a goodagent from an account-takeover bot, the screen-scraping-to-open-banking precedent, what Reg E and Reg Z do with "my agent did it," the managed-delegation posture (block anonymous agents, admit authenticated ones),and the two controls to turn up now, step-up authentication on anything that moves money and behavior baselining for the agent itself. Then gotchas, anti-patterns, and a Monday list. T Why this episode now OAuth has never been easier to learn, and MCP authorization guides exist, mostly written for developers building servers. This episode connects the two from the identity practitioner's chair: a six-minute OAuth refresher, MCP from zero, the handshake told as one scene (an intern walking up to room 412), what it changes about running an identity platform, the block-or-admit question for customer-facing agents, and why every change in the MCP spec this year was an OAuth lesson from years ago.
Thank you for listening LinkedIn: https://www.linkedin.com/in/rohit-agnihotri Email: [email protected]
More from The Identity Navigator
The Badge That Printed Itself: A SharePoint TokenForgery, Decoded Word by Word One line from my morning threat intel digest: “CVE-2026-55040. Critical. CVSS 9.1. SharePoint on-prem authentication bypass, under active exploitation.” This episode stops and takes the whole sentence apart, from first principles, until it reads likeplain English. Then we walk the heist: how a stranger with no account and no password handed SharePoint a homemade admin token, and SharePoint said welcome back.
Three labs. Five breached companies. One testing vendor named Irregular.
Everyone is using AI. Almost nobody can prove what it’s doing. In this solo episode, we design a runtime AI governance platform from scratch on a whiteboard, and then meet PulseAI, the product IAM veteran Amit Masand built at IDMEXPRESS. With the numbers justifying the problem, an honest competitive map, who it’s not for, and a free AI-traffic risk assessment for listeners. Links PulseAI by IDMEXPRESS: idmexpress.ai Company site: idmexpress.com) Amit Masand on LinkedIn: linkedin.com/in/amit-m-67b4692 Contact: [email protected] Listener offer: mention this episode for a complimentary risk assessment of your AI traffic, no commitment.
Previously on The Identity Navigator (recommendedlistening before this episode): Episode 40: SCIM - The protocol that finally made provisioning grow up
0-40% of SaaS accounts in the mid-2000s belonged to people who'd already left the company. Ghost accounts, quietly sitting there — and that sadly hasn't changed much today. This episode: where SCIM came from, what it actually solves, and what it still doesn't. Core RFCs: RFC 7642 — SCIM: Definitions, Overview, Concepts, and Requirements RFC 7643 — SCIM: Core Schema RFC 7644 — SCIM: Protocol