How to attack AI systems (and how to defend them) !!!!
summary
Shows concrete attacks and concrete mitigations: data poisoning of public crawls that flips dog images into “fish,” physical adversarial patterns (printed dots on glasses, four stickers that hide stop signs) that fool vision models at inference, and jailbreaks/prompt injections that extract harmful instructions. Uses FGSM and public web‑crawl datasets as attack tools, and defends with guardrails, privileged vs quarantine LLMs, adaptive guardrails and mandatory human‑in‑the‑loop checks rather than vague promises.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:51AI dual-use explained
- 5:26ML pipeline attacks
- 8:15Adversarial examples
- 13:55Jailbreak and prompt injection
- 20:44Prompt injection in agents
- 24:19Defending AI systems
- 28:00AI security trade-offs
- 33:42AI usage precautions
- 36:54AI security updates
- 41:32Q&A
related talks
44mHacking AI - How to Survive the AI Uprising
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.
46mA Short Introduction to Hacking Artificial Intelligence
37mKeynote - Hacking AI How to Survive the AI Uprising
21mAdversarial Attacks
53mHow to Break AI Systems (Before Someone Else Does)
37mHow to Break AI Systems (Before Someone Else Does)