Keynote: Towards Building Safe and Secure Agentic AI
summary
Automatic end-to-end red teaming frameworks generate exploitable inputs against deployed agents: agent visual uses iterative blackbox optimization to craft malicious website reviews that trigger indirect prompt injections and data exfiltration, while agent exploit analyzes source code to map attack paths (untrusted GitHub issues, browser inputs) and inject multi-point payloads that can run bash/python to kill execution servers. An open platform agent beats and a $1M+ competition aim for standardized benchmarks; a cyber gym evaluation finds models rising from ~10% to 28% success (single trial) and 67% with 30 trials, and agents discover 35 zero‑day vulnerabilities.
AI-written summary · about
chapters (11)
- 0:00Safe and secure agentic AI
- 1:23Security challenges in agentic AI
- 3:11Agent design and attack surface
- 5:45Red teaming for blackbox agents
- 10:07Red teaming for whitebox agents
- 13:42AgentBench evaluation
- 16:10Defense in depth for agents
- 19:47AI’s cybersecurity impact
- 20:30CyberGym benchmark
- 24:00Summary
- 26:27Closing remarks
related talks
19mKeynote: Towards Building Safe and Secure Agentic AI
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
53mHow to Break AI Systems (Before Someone Else Does)
29mHacking With AI and Hacking AI: Two Sides of the Same Problem
42mSecuring AI Agents...
13mKeynote: Agentic AI Is Running Your Infrastructure
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
20mAI Red Teaming and Automation: Exploring Societal Risks in...