Shadow AI & Sandbox Escapes: Why You Need an Agentic Control Plane?
Summary
Host Ashish joins Michael Leland from Aylain to discuss the "agentic control plane," a unified security framework for managing AI sprawl across browsers, desktops, and MCP gateways. They address risks like shadow AI, malicious skills, and goal-oriented agents that bypass guardrails, emphasizing visibility and token brokering. The episode highlights strategies for safe AI adoption, including model steering, non-human identity management, and calculating ROI through productivity gains and cost efficiency.
Topics discussed
Intro: AI sprawl, malicious skills, and the need for safe adoption
Guest intro: Michael Leland's background in cybersecurity and AI
Defining the Agentic Control Plane across browsers, desktops, and MCPs
Gaps in legacy security tools and the need for unified policy enforcement
Agent guardrails: Token brokering, model steering, and sandbox escapes
Threats from unvetted skills, plugins, and well-intentioned user errors
Visibility as the first win: Assessing shadow AI and risk scoring
Managing Non-Human Identities (NHIs) and the second-hop problem
Control plane architecture: From endpoint enforcement to cloud audit
Building trust in AI outputs via RAG, domain-specific models, and human-in-the-loop
Securing coding assistants like Claude Code and Cursor via desktop agents
Calculating ROI: Productivity gains, efficiency, and token cost optimization
You Laugh You Lose: Cybersecurity jokes and pickup lines
Outro: Island.io resources, demo offers, and podcast credits
Listen ad-free on Castria