Time Table

AI agents do more than respond to user prompts. They can autonomously read external documents, process tool outputs, and take actions based on the information they receive. In particular, with LLM-based AI agents, the boundary between instructions and data becomes increasingly blurred. Since attacks can also be delivered through natural language, traditional security approaches that rely on detecting known patterns or signatures are often insufficient.

As a result, organizations must address not only security risks associated with LLMs themselves but also threats that are unique to AI agents. These include prompt injection attacks that can lead to the disclosure of credentials or sensitive information, as well as the execution of unintended commands.

In this session, we will examine the evolving threat landscape for AI agents through real-world vulnerability examples and discuss defense strategies with a focus on boundary protection as one layer of a multi-layered security approach. Effective protection for LLMs and AI agents requires more than signature-based detection. It demands contextual understanding and risk-aware decision making.

In addition, organizations need visibility and governance not only over prompts and interactions with AI systems, but also over unauthorized AI usage, often referred to as “Shadow AI.” We will discuss how organizations can balance security and usability by minimizing false positives while enabling the safe adoption of AI agents. The session will also introduce practical experiences and initiatives from our organization in implementing governance and security controls for AI-driven environments.

Speakers:
Isao Takaesu Fellow, AI & Emerging Tech Acceleration Division
Shohei Koda AI security engineer, AI & Emerging Tech Acceleration Division