Laying the Groundwork for Cybersecurity Autonomy: Training Environments and Structured Challenges
| Authors | |
|---|---|
| Year of publication | 2026 |
| Type | Chapter in a book |
| MU Faculty or unit | |
| Citation | |
| Attached files | |
| Description | Recent advances in artificial intelligence have renewed interest in developing autonomous cybersecurity systems. However, despite significant progress in other domains, meaningful autonomy – whether in offensive or defensive operations – remains elusive. Even state-of-the-art large language models, initially heralded as a means to address the complexity of the cybersecurity domain without frequent retraining, underperform in both supervised tasks and unsupervised complex behaviors. In this paper, we argue that the development of domain-specific training environments and diverse scenario generation approaches is critical for advancing cybersecurity autonomy. We present a comprehensive overview of the current cybersecurity training platforms and evaluate their capacity to support the development of autonomous agents. Furthermore, we propose that structured cybersecurity challenges can serve as effective benchmarks for assessing both the current capabilities of AI-based systems and the complexity of tasks in the domain that is being considered. To illustrate this, we analyze recent challenges, including results from a cybersecurity competition organized by the authors, and discuss the implications for the future of autonomous cybersecurity systems and challenges |
| Related projects: |