As enterprises transition from simple chat assistants to autonomous agentic workflows, they face a critical security paradigm shift: the separation of "the brain" (cognitive LLMs) from "the hands" (execution environments). In this session, we will dissect the emerging security risks of the "agent skills" ecosystem—modular, reusable bundles of prompt context and executable scripts.

While frameworks have made agent building highly approachable, they introduce severe vulnerabilities. If an agent executes a poorly sandboxed script, a single prompt injection can compromise database credentials, mutate external systems, or trigger malicious system commands. We will explore a modern, secure architecture that solves "the excess agency" problem. Attendees will learn how to design a zero-trust runtime that isolates execution sandboxes, enforces principle-of-least-privilege on a per-tool/per-skill basis, and safely implements model-agnostic control planes in heavily regulated enterprise environments.

Key Learning Objectives for Attendees:
Understand the Anatomy of the Agent Harness: Learn how to isolate the agent runtime control plane from the untrusted sandboxes where model-generated code actually executes, ensuring sensitive credentials are never reachable by the model's environment.

Implement Skill-Scoped Security Policies: Discover how to declare network and file-system dependencies in declarative manifests (such as the open agentskills.io standard) to prevent high-trust database tokens from leaking into low-trust API integrations.

Mitigate Enterprise Scale Multi-Tenancy Risks: Analyze the structural, performance, and security trade-offs of always-on containers versus dynamic, on-demand execution sandboxes.