Advancing AI Security with Insights from the World’s Largest AI Red Team
summary
Lakera launches Gandalf.lakera.ai, an always-on red team game that accumulates 1 million players, 40 million interactions and about 25 years of aggregated hacking time across 24 levels; its data yields a prompt-attack taxonomy of 15 attack categories (obfuscation, partial/sequenced extraction, role‑play/DAN jailbreaks) that translate across foundation models, show model updates suddenly opening new attack vectors, and feed defenses used in Lakera’s enterprise product and a live RSA challenge with evolving levels and leaderboard prizes.
AI-written summary · about
chapters (8)
- 0:00The internet of agents
- 3:35Rethinking security for AI
- 7:39Introduction to Gandalf
- 13:58Insights from a global red team
- 17:03Taxonomy of AI attacks
- 26:29Evolution of threats
- 30:00RSA challenge and conclusion
- 33:24Q&A session
related talks
34mLLM Security 101 - An Introduction to AI Red Teaming
43mSecurity of GenAI & LLMs: A New Era of Ethical Hacking
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
20mAI Red Teaming Agent: Azure AI Foundry
41mHow Microsoft Approaches AI Red Teaming
57mAI as the New Hacker Frontier
25mProtecting AI Infrastructure
30mNavigating the New AI Security Landscape