Securing AI Agents that Handle Sensitive Data

AI agents can read data, call tools, and take actions on their own. That creates new paths for sensitive data to be manipulated or stolen, especially in regulated industries such as healthcare.

In industry

• Threat modeling for agent integrations. I have served as lead security architect for an AI-agent integration layer built on the Model Context Protocol (MCP), authoring a threat model that covered indirect prompt injection through tool inputs, agent tool abuse, unauthorized data exfiltration, and output-sanitization failures, and driving mitigations into the product before launch.
• LLM-assisted security tooling. I have built tools that use LLMs to speed up security review and incident investigation.
• Foundation-model security. I now work on securing foundation models on Amazon’s AI & AGI Security team.

The common thread with my earlier research on IoT Notary is accountability: being able to show what a system accessed, under whose authority, and under which policy.

This page describes my role at the level of my public CV; it contains no confidential employer information.