I'm about seven weeks late posting this, which for a guy who talks about the shrinking exploit window is a little embarrassing. Back in September I went back on GoTo's Practically Tomorrow podcast, this time with Mike Gutierrez, Senior Director of ITSG Strategy and Solutions Consulting at GoTo. My first time on the show was the Zero Trust episode in April 2025. This one was about AI, and how it's tearing up the security playbook.
Usual disclosure: I work at Zscaler, so yes, I have opinions about Zero Trust, and yes, I'm biased. You've been warned.
Agentic AI Is a Governance Problem First
The biggest theme of the conversation: AI tools don't wait for permission. Somebody connects an agent to the CRM on a Tuesday, three teams copy the idea by Friday, and by the end of the month nobody can tell you how many agents you have or what they can touch.
My argument is that uncontrolled AI tools need the same treatment as any other new kind of user:
- Compliance checks before an agent gets access to real data, not after the audit finds it.
- Red teaming the agent itself, because "it seemed fine in the demo" isn't a security test.
- Identity-based oversight, so every agent has an identity, an owner, and permissions you can actually revoke.
If you can't answer "who owns this agent and what can it reach," you don't have AI governance. You have AI vibes.
Zero Trust: A Network of One
We talked about Zero Trust as a "network of one." Every user, every device, every agent gets connected to exactly the app it's allowed to use and nothing else. No flat network to wander around in. If it's reachable, it's breachable, so stop making things reachable.
That idea matters more with agents, not less. A human attacker gets tired. An agent doesn't. Shrink what it can see and you shrink what it can break.
The Exploit Window Is Collapsing
AI-driven vulnerability discovery is making the gap between "a bug is disclosed" and "a bug is exploited" shorter than it's ever been. That changes the math on patch cycles. If your process assumes you have a few weeks after an advisory drops, your process is already out of date. Reduce exposure first, patch fast second, and assume the attacker read the advisory before you did.
The Boring Stuff Still Wins
For all the AI talk, some of the most important advice was the unglamorous stuff people keep skipping: MFA everywhere, and federated identity. The fanciest AI-powered attack in the world still loves a shared password on an app nobody tied into SSO.
For the MSPs Listening
A lot of GoTo's audience is MSPs, so we spent some time there:
- Take risk to the executives. Don't bury AI risk in a ticket. Frame it as a business conversation.
- Phase it. Big-bang security projects stall. Clear phases with clear wins get funded.
- Use AI, but check its work. AI is powerful, and it also hallucinates and can run up costs you didn't plan for. Keep a human spot-checking the output.
Bottom Line
AI didn't rewrite the fundamentals: know your identities, limit what's reachable, verify everything. It just made the cost of ignoring them show up a lot faster. Give the episode a listen, it's about 23 minutes, and thanks to Mike and the GoTo team for having me back.
Listen: Navigating Cybersecurity and AI, with Brian Deitch, on Practically Tomorrow