Hacking AI Systems: How to (Still) Trick Artificial Intelligence
summary
Survey of adversarial ML shows concrete exploits and defenses: training-set extraction reproduces images (example: Carlini reconstruction of training images from diffusion models), physical adversarial patches cause 3D-printed turtle misclassified as rifle, prompt leakage and RAG exfiltration let attackers retrieve secrets, and a meeting-invite payload demonstrates agent hijacking of Google Home/Gemini. Papers prove undetectable model backdoors exist. Practical mitigation emphasizes input/output sanitization, model segregation, adversarial training and continuous red‑teaming; guardrails slow but do not stop determined attackers.
AI-written summary · about
chapters (6)
- 0:00Intro
- 2:14Stepping into the shoes of an attacker
- 8:27Potential weaknesses in a model
- 21:44Exploiting models
- 32:10How to protect systems (quick primer)
- 35:03Outro
related talks
46mA Short Introduction to Hacking Artificial Intelligence
44mHacking AI - How to Survive the AI Uprising
45mSecuring AI: There Is No Try, Only Do!
42mHow to attack AI systems (and how to defend them) !!!!
44mComputers are Stupid: Protecting "AI" from Itself
52mSleeping With One AI Open: An Introduction to Attacks Against...
37mKeynote - Hacking AI How to Survive the AI Uprising
53mSecurity of Machine Learning Systems