ScenarioForge turns written requirements into validated cyber-range exercises
The arXiv paper describes a pipeline that builds and deploys capture-the-flag scenarios in the CORE emulator, then checks each step runs as intended.
In plain words
Building cyber-range exercises, where defenders practice on simulated networks, usually means manually wiring infrastructure, vulnerable services and success conditions. ScenarioForge is a system that converts high-level requirements into network topologies, services and challenge chains, then deploys them in the CORE emulator and verifies each objective actually works. The authors report all 226 generated scenarios in their dataset passed deployment and runtime validation.
What was announced
The authors say large language models can draft plausible exercise descriptions, but a convincing write-up does not guarantee deployable software or verifiable objectives. ScenarioForge addresses this by grounding generation in actual resources: it produces topologies, services, traffic, segmentation and challenge chains, then instantiates them in CORE and validates them. The accompanying dataset holds 226 materialized scenarios built from 294 Vulhub images and 85 custom challenge-resource generators, whose 867 individually verifiable steps split into 457 vulnerability-backed and 410 generated-resource-backed instances. Every scenario passed construction, preview, sequencing, deployment, strict runtime validation and live artifact checks.
In a separate 50-prompt matched benchmark, the authors report that ScenarioForge paired with Qwen3.8 or GPT5.6 achieved 50 of 50 topology and build successes and 160 of 160 intent fields, while direct frontier generation reached 41 of 50 builds and 147 of 160 fields, and only when given the full cumulative scaffold of structure, vocabulary, policy and container-image guidance. They note the benchmark omitted ScenarioForge's downstream challenge-resource generation, dependency and flag sequencing, deployment and live validation, so the comparison covers only part of the pipeline. The artifacts are presented as supporting reproducible exercises for training and evaluation.
What it does not show
- Evidence is limited to the paper's abstract; the full methodology, experimental detail and peer review status are not available in the supplied passages.
- The benchmark comparison and validation results are self-reported by the authors and have not been independently confirmed.
- The 50-prompt benchmark excluded challenge-resource generation, sequencing, deployment and live validation, so it does not measure the full pipeline end to end.
For security and IT teams
- Scope
- The reported scenarios run in the Common Open Research Emulator (CORE) and rely on Vulhub images plus custom challenge-resource generators, so coverage is bounded by those resources.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.