Keynote: Towards Building Safe and Secure Agentic AI
summary
Describes an agent visual blackbox red‑teaming method that uses iterative optimization to craft environment payloads (example: a malicious review that triggers indirect prompt injection), a standardized agent beats platform with a $1M competition for reproducible benchmarks, and a defense stack recommendation (layered defenses, least privilege, runtime monitoring) plus a programmable guardrail system called proent. Reports empirical results: cloud sonet 4 finds ~17% known vuln, GPT‑5 ~22%, cloud sol 4.5 ~28%, and agents discover 35 zero‑days.
AI-written summary · about
chapters (10)
- 0:00Safe and secure agentic AI
- 1:03Security challenges in agentic AI
- 3:51Attack surface in agentic systems
- 5:36Evaluation and risk assessment
- 8:51AgentBench for agent evaluation
- 10:42Layered defense strategies
- 12:50Frontier AI and cyber security
- 15:43Frontier AI Cyber Security Observatory
- 16:47Shifting AI power to defenders
- 18:12Resources and next steps
related talks
27mKeynote: Towards Building Safe and Secure Agentic AI
53mHow to Break AI Systems (Before Someone Else Does)
01h 11mKeynote - Securing AI: Building Trust in the Era of AI
13mKeynote: Agentic AI Is Running Your Infrastructure
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
01h 01mKeynote: Guardians of the AI Era: Navigating the Cybersecurity Landscape of Tomorrow
37mKeynote - Hacking AI How to Survive the AI Uprising
42mSecuring AI Agents...