Episode 236 - GluuFederation/identerati-office-hours GitHub Wiki

Title: Who Let the Agents In? IGA and the Agent Gateway

• Host: Mike Schwartz, Founder/CEO Gluu
• Guest: Nov Rosenman, IT Infrastructure & Cloud Engineer at Monday.com

Channels

• Linkedin Event
• Youtube Video

Description

AI agents use MCP tools to take actions, but broad credentials and prompt injection can turn those actions into security risks. We’ll explore how an MCP gateway identifies the agent and the person it acts for, authorizes each tool call, protects credentials, and records what happened. We’ll also cover tool tampering, injection checks, human approval for high-risk actions, and where to deploy the gateway.

Homework

• MCP specification 2026-07-28, Authorization Read for: resource-bound tokens, the ban on passing tokens through, the stateless transport and per-request headers that make gateways practical, the client-credentials and Enterprise-Managed Authorization extensions.

• NSA Cybersecurity Information Sheet, "Model Context Protocol (MCP): Security Design Considerations for AI-Driven Automation". Released on May 20, 2026 by the NSA's Artificial Intelligence Security Center, it is the clearest statement of the agent-specific risks: trust boundaries, dynamic tool invocation, implicit trust, token lifecycle, human approval.

• OWASP MCP Top 10. Read for: tool poisoning, scope creep, token mismanagement, intent-flow subversion, shadow servers, missing audit. The talk maps each gateway function to the items it closes.

Takeaways

• ⚡
• ⚡
• ⚡
• ⚡

Livestream Archive

• Audio
• Transcript