Adversarial AI: Why it Matters Now
summary
DHS S&T treats adversarial AI as an immediate, practical threat distinct from traditional cyber, focusing on concrete vectors: evasion attacks that use adversarial perturbations (glasses that make Reese Witherspoon read as Russell Crowe, sticky notes that fool stop signs for self‑driving cars), purchasable anti‑facial‑recognition clothing, deepfake Zoom fraud that convinced a clerk to wire $25M, morphing IDs that match two people, LLM jailbreaks that disclose instructions (hotwiring example), model theft and AI‑designed control attacks analogous to Stuxnet.
AI-written summary · about
chapters (9)
- 0:00Introduction to adversarial AI
- 1:13Defining adversarial AI scope
- 3:01AI-based attacks on AI systems
- 5:52Key adversarial AI threats today
- 8:26Evasion attacks explained
- 10:52Generative deceptive AI risks
- 14:32Morphing and LLM misuse
- 17:00AI and cybersecurity evolution
- 20:35Future threats and conclusion
related talks
47mWhat AI Security means for you as a developer: Time-focused approach
45mHardening AI/ML Systems - The Next Frontier of Cybersecurity
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
46mA Short Introduction to Hacking Artificial Intelligence
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.
37mFundamentals of AI safety and security
49mThe Intersection of AI and Cybersecurity
21mAI Deception: Fooling (Artificial) Intelligence Is Easier than You Think