Event-driven temporal graph networks can bridge the simulation-to-reality gap in multi-agent cyber defense by processing asynchronous, noisy alerts in continuous time rather than synchronous ticks, enabling policies trained in simulation to work on real systems.
NetForge_RL is a cyber defense simulator that trains AI agents to protect networks in realistic, continuous-time conditions rather than simplified turn-based games. It uses a new technique called CT-GMARL that processes irregular security alerts like a human analyst would, achieving 2x better performance than existing methods and successfully transferring trained policies to real systems.