Terra Security’s new AI Native Cybersecurity Podcast opens with a conversation that security teams can put to use right away. Hosted by Terra co-founder and CEO Shahar Peled, the series presents field notes from security teams defending against and operating like AI adversaries. Its first guest is CJ Moses, CISO of Amazon, who leads security engineering and operations across the company and has worked at Amazon since December 2007.

Here is what the episode covers and why each point matters.

Exposure Windows Now Run in Minutes

Moses identified speed as the defining change of the generative AI era: “The speed at which you need to respond to them.” Patching cycles of 30, 60, or 90 days were the default mindset as recently as six months ago. Today, adversaries using Mythos and similar tools can find and exploit vulnerabilities in seconds.

Amazon’s honeypot infrastructure, MadPot, recorded one attacker completing intrusion, reconnaissance, and lateral movement across every reachable system in 12 minutes and 42 seconds. That kind of activity once took weeks or months.

Why it matters: Response plans built around monthly cycles need to be rebuilt around minutes.

Adversaries Are Scaling Insider Operations With AI

North Korean operators are using AI in ways that surprised even seasoned defenders. Moses said one “normally wouldn’t think of the DPRK as being frontline AI innovators or users, but they’re actually using AI to generate personas, backstopping entire identities.” The goal is to get hired at companies like Amazon, which has stopped at least 2,900 DPRK actors in the hiring process.

Moses noted that culture is the main reason for Amazon’s in-office approach, but security benefits too: “it’s a lot easier to know that you’re showing up for work that it’s actually you than a remote worker that you really don’t know.”

Why it matters: Identity verification in hiring is now a frontline security function.

The Shared Responsibility Model Still Holds

Moses wrote the AWS shared responsibility model in 2008, and it continues to apply to AI workloads. With a service like Bedrock, AWS secures the infrastructure and supplies certain guardrails, while the customer owns everything above the handoff point. His guiding rule: whoever has access carries responsibility.

Why it matters: Teams deploying AI agents can use a framework they already understand.

Faster Pipelines Are Safer Pipelines

Machine-written code is no exception to the rules. “AI generated code still has vulnerabilities just the same way human code does,” Moses said, since models learn from human code. Amazon’s answer is to run CI/CD faster, scanning code inline during each iteration, then completing a final scan and penetration test. AI-assisted patching has multiplied coverage, and MadPot intelligence guides prioritization.

Why it matters: Slowing development to add review time extends exposure rather than reducing it.

Autonomy Depends on the Blast Radius

Amazon separates decisions into two-way doors and one-way doors. For the first kind, Moses asked: “Anything that you can back out of rather easily or roll back quickly without any impact, why not move forward?” For one-way doors, such as an AI chaining its way toward root access or customer data, the system should stop, explain the potential impact, and ask a human.

Why it matters: Guardrails defined in advance let teams gain speed without sacrificing control.

Time Is the Metric That Counts

Moses closed with a forecast for the year ahead: “The measure of a good security program won’t be how many vulnerabilities you found.” Mean time to defense, the gap between finding a vulnerability and patching, blocking, or isolating it, will become the number that matters most.

Why it matters: Reporting should shift from volume found to exposure closed.

More From the Conversation

The episode also covers Moses’s background, which includes leading intrusion analysis at the FBI’s Cyber Division, serving as a Special Agent with the Air Force Office of Special Investigations, and competing as an SRO GT America GT2 race car driver. He draws a parallel between racing and AI adoption, describing how trust in a car is earned lap by lap before going flat out.

Peled also shares how Terra applies its own version of shared responsibility to TerraTorch, which places an AI hacker in customers’ hands under a scope they define.

How to Listen

The full episode is available at ainativecybersecurity.com, and more about Terra Security is at terra.security. Peled has invited listeners to send their questions his way and has said he will pass them along to Moses, who plans to return as the field keeps evolving.