Safety Testing of the AI Agent: Vulnerabilities and Attacks Beyond the Chatbot

This conference talk from MLOps World: Machine Learning in Production features Alexander Borodetskiy, VP of Growth, AI Safety at Toloka AI, exploring the critical challenges of safety testing AI agents that go beyond simple chatbots. Discover a new safety evaluation framework that reveals how AI agents can be compromised through specially crafted web pages, files, and OS environments. Learn about vulnerability testing methods, including how prompt injections can lead to sensitive data leaks, techniques for organizing comprehensive safety testing, and potential risks of agent hijacking for malicious purposes. This 19-minute technical presentation provides valuable insights for developers and security professionals working to build more robust and secure AI systems as agents evolve to browse the web and interact with computer systems.