In this episode, John Cranmer is joined by Gil Elbaz, Co-Founder and Chief AI Officer at Onyx Security, the company building the secure AI control plane for the agentic era. Gil, formerly in the CTO office at Nvidia working on multi-agent infrastructure, argues that the AI attack surface has fundamentally changed. The old boundary between what is yours and what is not has collapsed, and every agent, every MCP, every tool connection is now part of the perimeter.

The conversation moves from the rise of shadow AI inside enterprises, where employees reach for any tool that makes work faster, to the harder problems underneath: rogue agents taking destructive actions they were never asked to take, multi-agent communication happening invisibly through Slack and email rather than direct A2A channels, and offensive AI systems like Mythos finding more zero-days in weeks than internal red teams find in years. Gil shares his playbook for getting your arms around it: visibility first, then policy, then runtime protection. As he puts it, the right seatbelt is what lets you drive faster.