Battling the Machine: Cutting-Edge Attacks and Defenses in Generative AI
summary
Examples show GPT‑3/GPT‑3.5 yielding two real Windows 10 activation keys exposed after web-sourced training data and a jailbreak that produced a poison recipe; instruction‑hijack prompts like “ignore the above” override templates; backdoor triggers flip an “alarm” intent into “email”; model extraction clones API behavior via query‑response distillation. Defenses tested include input sanitization plus a logistic‑regression prompt classifier, poisoned‑example robust training, watermarking fingerprints, federated and encrypted training.
AI-written summary · about
chapters (10)
- 0:00Generative AI security threats
- 1:08LLM vulnerabilities and prompt attacks
- 4:21Prompt injection techniques
- 10:53Data poisoning attacks
- 15:51Model extraction via API
- 18:31Defending with prompt guarding
- 21:31Robust training and data filtering
- 23:22Model watermarking for theft
- 25:28Future AI security directions
- 29:03Summary
related talks
49mChatGPT: A New Generation of Dynamic Machine Based Attacks?
37mHow to Break AI Systems (Before Someone Else Does)
32mSecurity Implications of Artificial Intelligence Synthesizers
53mHow to Break AI Systems (Before Someone Else Does)
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
25mAttack, Defense & Danger in the Age of AI
36mHow NOT to Train Your Hack Bot: Dos and Don'ts of Building Offensive GPTs
43mWords as weapons: The dark arts of Prompt Engineering