Page 275 - Cyber Defense eMagazine September 2025

sacrificing performance – a balance I’ve helped numerous financial institutions achieve. This article
examines the Defense in Depth approach to securing AI applications, covering data protection, model
security, and enterprise integration.

Modern AI workloads face critical security risks including:

• Prompt injection attacks – a threat I’ve seen increasingly targeted at chatbots
• Model output manipulation leading to hallucinations challenging for critical applications
• Training data poisoning attempts
• Model theft through extraction attacks that is a growing concern for proprietary trading models

Drawing from my recent experience heling a G-SIB secure their generative AI infrastructure,
organizations follow NISTs Defense in Depth (DiD) methodology to protect assets across all layers- from
infrastructure to model endpoints. What’s particularly effective about this multi-layered strategy is how it
combines robust security controls with native safeguards embedded within AWS Generative AI services.

From my work I’ve learned that organizations can confidently advance their AI initiatives while upholding
stringent standards for data protection, compliance, and operational security by following a methodical
approach tailored to their risk profile.

AWS Defense in Depth Strategy for Security

While Defense in Depth (DiD) isn’t a new concept, its application to AI workloads requires specific
considerations. Let me share a real example: When implementing Bedrock for a major investment bank,
we discovered that traditional DiD approaches needed significant adaptation to handle the unique
challenges of foundation models.

Defense in Depth (DiD) employs multiple layers of security controls to protect data, applications, and
infrastructure. What makes this especially crucial for AI workloads is the dynamic nature of model
interactions. For instance, one of my clients discovered that traditional data loss prevention (DLP)
solutions were insufficient for catching sensitive data in model outputs - we had to implement additional
semantic analysis layers.

Each security layer provides distinct but complementary controls. Drawing from my experience here’s a
breakdown of how these layers work together in practice:

Security Layer Primary Function Key AWS Security Services

Data Protection Protect data at rest and in transit AWS KMS

(Core Layer) AWS CloudHSM

Amazon Macie
AWS Secrets Manager

Cyber Defense eMagazine – September 2025 Edition 275
Copyright © 2025, Cyber Defense Magazine. All rights reserved worldwide.