Ghost in the Machine - Exploiting and Securing AI Agent Systems

Explore AI security vulnerabilities and defense strategies in this 48-minute conference talk that examines real-world attack vectors targeting AI agent systems. Analyze actual security breaches from major organizations to understand how adversaries exploit AI systems in ways traditional security measures cannot anticipate. Dive into the technical mechanics of advanced attacks currently compromising production systems, including data poisoning that subtly manipulates agent decisions, adversarial inputs that bypass detection systems, and stealthy backdoors that undermine model integrity. Learn from the safety approaches used by leading AI labs like OpenAI and Anthropic, and discover open-source tools for securing locally deployed AI agents. Master actionable defense blueprints with concrete countermeasures for every attack vector discussed, from hardening training pipelines to implementing real-time input validation. Walk through proven methodologies for auditing AI systems and integrating security checks throughout the development lifecycle, complete with practical implementation examples. Map the evolving threat landscape including emerging risks like AI-powered social engineering and defenses such as automated vulnerability scanning. Gain the knowledge to identify and mitigate current AI agent attack vectors, apply security assessment frameworks tailored for AI systems, design secure AI agent architectures using open-source tools, implement continuous monitoring for model integrity, and integrate AI-specific security practices into existing development workflows.