Keynote:Bend the Machine: The Hacker Path to Autonomy in the Real World
summary
DARPA's Cyber Grand Challenge builds an autonomous cyber reasoning system that must find exploits, produce patches, and avoid downtime; the winning system used symbolic execution plus fuzzing but the real edge was a robust automated testing engine. Real-world adoption breaks on integration (no Docker, PowerPC/embedded binaries), differing notions of safety, and customer value until an 80% code coverage test suite convinced procurement. Reinforcement learning on CTFs trains agents (Claude hits top 3% PICOCTF, 19/20 Airbnb) but fails at Plaid/Defcon, so reinforcement learning shows promise but not yet trustable autonomy.
AI-written summary · about
chapters (19)
- 0:00AI and hacking trust
- 1:51George Hotz on hacking
- 3:15AI in security question
- 5:04Autonomy loop in AI security
- 7:56DARPA Cyber Grand Challenge
- 11:05Mayhem’s cyber reasoning
- 15:43Mayhem wins challenge
- 16:26Testing key to autonomy
- 18:11AI vs world’s hackers
- 20:39AI adoption challenges
- 23:09Customer value over tech
- 27:49AI adoption loop
- 28:54Human improvement practice
- 31:01Capture the Flag training
- 32:28AI training with reinforcement
- 34:59Claude in hacking contests
- 36:29Detecting AI failures
- 39:01Three loops for AI trust
- 40:29Making failure visible
related talks
01h 01mKeynote: Guardians of the AI Era: Navigating the Cybersecurity Landscape of Tomorrow
27mKeynote: Towards Building Safe and Secure Agentic AI
35mThe long road to autonomous cybersecurity
19mKeynote: Towards Building Safe and Secure Agentic AI
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
23mCode Security Reinvented Navigating the era of AI
49mAutonomous Hacking Systems - Future Risk or FUD?
36mKeynote: The Future of Software