Vybhavi_08

AgentTrace is a security-testing prototype that explores what happens when a relatively robust AI model is integrated into a larger application. Rather than focusing only on model jailbreaks, AgentTrace examines whether an AI-powered system can complete a legitimate user request while unintentionally performing an unauthorized action. Built using Meta Muse Spark 1.3, the project demonstrates this through a simulated support-ticket application. A user requests a ticket summary, but an insecure application integration also updates an unrelated ticket without authorization. AgentTrace compares four configurations: a model-only baseline, an insecure application integration, additional model-level security instructions, and an application-level authorization gate. The experiment shows that stronger model instructions alone do not necessarily prevent insecure application behavior, while a deterministic authorization check can block the unauthorized update without interrupting the legitimate task. Future work will explore more realistic scenarios, interconnected AI systems, multi-step workflows, and how small unauthorized actions could propagate into larger consequences.