When an AI agent is given a tool interface to read, write, and execute code within an enterprise pipeline, it becomes a high-value target for emerging threat vectors. If an untrusted input introduces a prompt injection, or if an agent autonomously processes a malicious pull request, the security blast radius can instantly compromise your engineering environment.
This session breaks down the security architecture of a multi-agent software factory using Model Context Protocol (MCP) and Azure OpenAI. We will demonstrate practical attack scenarios—including indirect prompt injection and excessive permission exploitation—to illustrate exactly how autonomous agents can be manipulated. Moving beyond the threats, we will design a robust, zero-trust defensive framework. You will learn how to build secure tool execution sandboxes, implement runtime monitoring, and establish automated guardrails to intercept malicious agent behavior before it touches production infrastructure.
Participant Takeaways:

- Threat Modeling Agents: Identifying vulnerabilities in multi-agent architectures, focusing on prompt injection and tool-abuse vectors.
- Secure MCP Architecture: Engineering isolated, least-privilege tool execution environments for Model Context Protocol integrations.
- Agent Observability: Implementing logging, auditing, and telemetry to detect and alert on anomalous AI system behaviors in real time.