AI Security Foundation

How can we make AI systems useful while limiting what an attacker, an unreliable model, or an overpowered agent can make them do? Students will attack, defend, and evaluate AI systems, moving the boundary of trust outward ring by ring. Topics include threat modeling and trust boundaries; adversarial examples; data poisoning, backdoors, and the model supply chain; memorization, privacy attacks, and defenses; prompt injection, jailbreaks, retrieval and guardrail security; securing AI agents with least privilege, sandboxing, and tool authorization; misuse, deepfakes, watermarking, and provenance; and confidential computing and cryptographic private inference.