Crash Test for AI Agents: Security Check for Model, Agent and Harness

Tested live: how does an AI application respond to prompt injection, jailbreaks and data leakage? Stefania Kalinska demonstrates the security check with the ECODYNAMICS AI Security Agent.

An AI agent is only as secure as its model, its tools and the harness around it. Stefania Kalinska tests an AI application live on a dedicated test system, using the ECODYNAMICS AI Security Agent. It comes with 21 test templates and 17 test sets with 164 attack patterns and covers 7 of the 10 OWASP categories for LLM applications: from prompt injection and leakage of data and system prompts to excessive agency. A second, independent attacker model writes multi-turn attacks. You will see how an attack turns into a traceable finding with evidence and which safeguards follow from it. Testing only takes place on approved test systems. The session runs in parallel to the expert talks.