A new paper on arXiv introduces an Autonomous Cyber Operations (ACO) gym, a framework designed to train and evaluate AI models for cyber defence. The authors propose a unified environment that brings together several components needed for realistic testing: high-fidelity network emulation, models of benign user behaviour, and profiles of adversarial activity.
The gym is intended to address a common gap in cyber AI research, where models are often trained in simplified or static settings that do not reflect the dynamics of real networks. By simulating realistic traffic and attack patterns, the framework could provide a more reliable way to measure how well an autonomous defender performs before deployment.
Because the source is an abstract, the paper's specific implementation details and experimental results are not yet available. The contribution at this stage is the proposed architecture and the goal of standardising how autonomous cyber defence agents are developed and compared.