AI Safety in the Generative AI Era
summary
Claims a red teaming platform called virtual red with about 300 algorithms finds concrete vulnerabilities like privacy leakage and brand-risk exploits, cites the 2016 trading hack that wiped out ~$136 billion and a recent 60–70% stock swing from a typo, and shows no single model dominates safety. An open-source benchmark DecodingTrust on Hugging Face ranks models across fairness, hallucination, privacy and robustness. A hardened model using energy-based data augmentation plus a k-nearest neighbor layer yields near-zero jailbreak success versus OpenAI moderation API, Nemo and earlier LlamaGuard.
AI-written summary · about
chapters (8)
- 0:00Introduction to AI safety
- 0:42Real-world security risks
- 3:40A platform for AI safety
- 6:28Assessing model vulnerabilities
- 9:12Evaluating trust and risk
- 12:26Regulatory compliance
- 13:37Resilient model design
- 15:33Future outlook and conclusion
related talks
25mProtecting AI Infrastructure
50mSecurity as Part of Responsible AI: At Home or At Odds?
39mDefensible Application Security for the Artificial Intelligence Era
30mIntroduction to AI Safety. Let's be ready before it's too late!
24mHow to Build Trustworthy AI
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
01h 05mNavigating the Security and Privacy Landscape of Modern AI
27mKeynote: Towards Building Safe and Secure Agentic AI