Don't Get Hacked While You Hack - Detecting Prompt Injection in the Wild

Learn to detect and defend against prompt injection attacks in AI models and agents through this conference talk from Ekoparty Security Conference. Explore adversarial techniques that leverage embedding-comparison methods to identify prompt-injection vulnerabilities in real-world scenarios. Discover how these detection methods can reveal security weaknesses while also understanding how they might be exploited by malicious actors. Examine practical defense strategies including human-in-the-loop approaches to mitigate prompt injection risks. Gain insights from security researcher Martin Doyhenard's research-backed analysis of this emerging threat landscape, drawing from his extensive experience in exploiting HTTP servers, web applications, and his recent work on HTTP Response Smuggling and SAP Inter-Process Communication vulnerabilities.