Suyog Patil

Placeholder session for confirmed speakers without an assigned session, allowing them to be pushed to Swapcard and Visit.

PwnGraph is an open-source runtime attack graph engine designed to help security researchers and developers identify, understand, and visualize multi-step attack chains in AI agents. Traditional AI security tools often report isolated vulnerabilities such as prompt injection or insecure tool usage, but they rarely demonstrate how these weaknesses combine to compromise an autonomous agent.

During this live demonstration, attendees will see how PwnGraph attaches to a running AI agent, traces reasoning steps, monitors tool invocations, and executes adversarial test cases including prompt injection, indirect prompt injection, tool abuse, memory poisoning, and data exfiltration scenarios. As attacks execute, PwnGraph automatically builds an interactive attack graph that reveals the complete exploitation path from the initial entry point to the final impact.

The session will also demonstrate how findings are mapped to the OWASP LLM Top 10, enriched with security context, and exported as actionable reports for engineering teams. Attendees will learn how runtime attack graph analysis provides deeper visibility into AI agent behavior than traditional vulnerability scanning and how it can be integrated into AI security testing workflows.

Whether you're building AI agents, conducting security assessments, or performing AI red teaming, PwnGraph provides a practical approach to understanding and securing complex AI agent ecosystems.