Agentic AI Needs Humans on the Loop with Michael Chavira

Episode notes

As AI systems start acting on their own, the hardest security problem is no longer approving what a model outputs but governing the conditions under which it is allowed to act at all. Michael Chavira, co-founder of Axiologic Solutions, argues that most organizations confuse governance with security, writing policies and standing up AI councils that describe intent while leaving the enforcement layer, where a model actually touches data, underbuilt. His answer is to move humans out of the approve-every-action role and onto the loop, where they define the boundaries an agentic system runs inside and watch for it to step out. That control is only as good as the data underneath it, and for most enterprises, the data is a weak point. Chavira makes the case that AI security is ultimately data security, since a model is only as trustworthy as the data feeding it, and he walks through why shadow AI, unlabeled data and untracked workarounds quietly undermine even well-governed systems. For tools that answer the same question differently each time, he recommends harness engineering and continuous data operations as the way to secure the system rather than chase every output.

• Autonomous Agents Need a Verified Identity with Rosalyn Curato78 / 100
• AppSec Needs AI Employees, Not More Tools with Shan Kulkarni64 / 100
• Unleash AI or Block It? The Answer Is Both with Steven Walchek
• AI Agents Aren't People. Don't Give Them the Keys with Ido Shlomo
• Human Behavior Is the Attack Surface | Part 2 with Jeremiah Baker