Reinforcement Learning for Autonomous Cyber Defense
summary
Describes how recent open-source “cyber gyms” (CybORG, CyberBattleSim, Yawning Titan, FARLAND) let teams train reinforcement-learning defenders against scripted red agents and simulated users, with public CAGE competitions and reproducible YAML scenarios. Publication volume runs around 100 RL+cyber papers per year (Krystal’s repo search adds ~15% more). CAGE2 winner uses 36 actions 62 observations 6,000 parameters and environments give only negative rewards by default. Practical demos show agents learning over many iterations; expect toolchains like CybORG and CAGE competitions for hands-on evaluation.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:47The Cyber Challenge
- 3:45What is Autonomous Cyber Defense?
- 5:37Where Do Researchers Disagree?
- 10:29What is Reinforcement Learning?
- 14:15Timeline of Key Events
- 15:36Reinforcement Learning for ACYD
- 18:22CAGE2 Challenge
- 20:24Red Agent Kill Chain
- 32:36Technical Challenges
- 37:34Sizing up Different Models
- 39:58Development of Autonomy Systems
- 42:08RSAConference 2023
related talks
35mThe long road to autonomous cybersecurity
37mA practical perspective on training reinforcement learning bots.
47mHow we use Reinforcement Learning (...)
41mKeynote:Bend the Machine: The Hacker Path to Autonomy in the Real World
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
28mSAIDA RL: An Open-source Reinforcement Learning Platform
28mDeep Reinforcement Learning: theory, intuition, code
16mAdversarial AI Agents for Web3 Security