Guest:

• Nicholas Carlini, Research Scientist @ Google

Topics:

• What is your threat model for a large-scale AI system? How do you approach this problem? How do you rank the attacks?
• How do you judge if an attack is something to mitigate? How do you separate realistic from theoretical?
• Are there AI threats that were theoretical in 2020, but may become a daily occurrence in 2025?
• What are the threat-derived lessons for securing AI?
• Do we practice the same or different approaches for secure AI and reliable AI?
• How does relative lack of transparency in AI helps (or hurts?) attackers and defenders?

Resources:

• "Red Teaming AI Systems: The Path, the Prospect and the Perils" at RSA 2022
• "Killed by AI Much? A Rise of Non-deterministic Security!"
• Books on Adversarial ML