Winning by deception: tactics, techniques, and procedures of adversarial A.I.
summary
Four clear adversarial TTPs are defined: poisoning, evasion, extraction, inference with concrete examples — data-poisoned stop signs that change vehicle behavior, the Defcon subreddit API extraction and breast‑cancer model cloning, a reported 18–20% hallucination rate and a claimed 1.76 trillion‑parameter scale, and dark‑web tools like wormGPT and FraudGPT that automate phishing and malware. Practical defenses include MITRE ATLAS mapping, DARPA’s GARDA/GARD tools and apricot dataset testing, watermarking, rate‑limits and differential privacy; several tested bypasses still succeed.
AI-written summary · about
chapters (18)
- 0:00Introduction and speakers
- 1:31About the presenters
- 2:54AI basics and hallucinations
- 8:14Guardrails and bypass techniques
- 11:04Dark web AI tools
- 14:36Poisoning attacks explained
- 16:23AI evasion tactics
- 18:08Conversation overflow attack
- 19:10AI model extraction
- 20:54Inference attacks on AI
- 22:38ATLAS framework overview
- 24:26Countermeasures and hardening
- 26:34System-level AI security
- 27:59Evasion countermeasures
- 28:43Extraction and inference defenses
- 30:07Microsoft’s AI security measures
- 31:32Future threats and trends
- 35:23Open Q&A session
related talks
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
42mHow to attack AI systems (and how to defend them) !!!!
30mBattling the Machine: Cutting-Edge Attacks and Defenses in Generative AI
44mHacking AI - How to Survive the AI Uprising
52mSleeping With One AI Open: An Introduction to Attacks Against...
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
40mEvasion, Poisoning, Extraction, and Inference: Tools to Defend and Evaluate
22mAdversarial AI: Why it Matters Now