Send us Fan Mail

You cannot secure AI by filtering prompts alone. If an AI agent can retrieve sensitive data or execute actions, prompt injection becomes an identity, authorization, and architecture problem.

In this episode of Musings from the Cyber Trench, I sit down with Sundar Krishnamurthy, Senior Security Architect at Expedia Group, to discuss how organizations can secure AI in real enterprise environments.

We examine why prompt-level guardrails are not enough, how non-human identities should remain connected to the people they represent, and why security controls must be enforced close to the data and action being requested.

Sundar shares practical insights on:

• Turning security assumptions into automated tests
• Using LLMs to support code reviews and threat modeling
• How indirect prompts can bypass basic guardrails
• Enforcing authorization at the point of data retrieval
• Maintaining human and non-human identity context
• Isolating vector stores to reduce cross-tenant exposure
• Avoiding security shortcuts during MVP development
• Protecting private data when using AI models

The central message is simple: AI security cannot depend on trusting the model or recognizing every malicious prompt. Identity, authorization, data isolation, logging, and defense in depth must be built into the underlying architecture.

ABOUT SUNDAR KRISHNAMURTHY

Sundar Krishnamurthy is a Senior Security Architect at Expedia Group. He conducts architecture reviews across applications, cloud, networks, infrastructure, and AI and machine learning security. His background includes work with AWS, SAP Concur, and Microsoft.

Responsible for ICAM, Zero Trust, or identity security in a federal agency, prime, or large regulated enterprise?

If you’re trying to move from strategy to execution, start with Zephon’s Zero Trust Readiness Assessment: zephon.tech/zt

Questions or guest ideas? Email defend@zephon.tech