SANS Cyber Defense Initiative 2026 - SANS@Night: Rehearsing the End of the World: A MITRE-Led TTX on Protracted Cyber Conflict

Prompt injection is one of the most critical and misunderstood risks in modern AI systems, especially as applications evolve from simple chatbots to autonomous, agentic workflows.

This talk takes a deep dive into prompt injection by unpacking what prompts are, where they come from, and why untrusted instructions hidden in user input, external content, tools, and memory can fundamentally alter AI behavior. We will examine direct and indirect prompt injection, contrast chatbot risks with agent-specific risks, and explore how single-agent and multi-agent systems expand the attack surface and blast radius. Through real-world examples, common payload patterns, multi-hop injection scenarios, and the challenges introduced by long context windows, attendees will gain a practical understanding of how these attacks work in practice.

The session will close with OWASP-aligned mitigations and broader defensive strategies to help security practitioners, developers, and architects build more resilient AI applications.