A dataset of emulated APT-style cyberattack chains: end-to-end runnable, symbolic-planned, and openly modeled.
Aurora is a system for automatically constructing high-fidelity, end-to-end, APT-style attack chains for emulation. It combines classical (PDDL-based) planning with LLM-assisted knowledge extraction to compose real attack tools (Sliver, Meterpreter, Atomic Red Team, native Windows utilities) into causally coherent multi-step chains — the kind of chains you'd otherwise write by hand for a red-team exercise.
This repository is the output dataset of Aurora. Every chain here is represented in a shared symbolic model (AALM — Aurora Action Language Model), comes with an executable Python wizard, and is meant to be run against real emulation environments to produce authentic telemetry.
📄 Paper: From Sands to Mansions: Towards Automated Cyberattack Emulation with Classical Planning and Large Language Models (ACNS 2026) — arXiv preprint
🌐 Homepage: auroraattack.github.io
Aurora chain-005 running end-to-end on env0 (3 minutes). See Watch it run for more chain videos, or jump straight to the YouTube playlist.
The repo is organized into two independent axes: releases of the dataset and the environments those releases were built against.
Aurora-demos/
├── README.md ← you are here
├── LICENSE ← Apache 2.0
├── CITATION.cff
│
├── attacks/ ← chain dataset, versioned
│ └── v2.0/ ← the current release
│ ├── README.md ← what's in this release, how to run it
│ ├── chain-001/ ← one directory per attack chain
│ ├── ...
│ └── docs/
│
└── environments/ ← emulation testbeds, shared across releases
└── env0/ ← the Windows-Server-2022 testbed used for v2.0
├── README.md
├── quick-start/ ← one-click OVA path
└── from-scratch/ ← manual build path (transparent, hypervisor-agnostic)
New releases are added under attacks/vX.Y/ without touching prior releases (git tags freeze historical states). New testbeds are added under environments/envN/. Releases declare which environment(s) they were validated against in their own README.
| Release | Chains | Testbed | Highlights |
|---|---|---|---|
| v2.0 | 101 | env0 | Updated AALM ontology, msf-HTA / VBS / fodhelper action families, per-chain readmes with MITRE ATT&CK coverage |
Legacy 1.0 chains(250 CTI-derived chains from the original release)are preserved on the v1.0 git tag — they use an older AALM predicate scheme and are not maintained forward.
| Name | Description |
|---|---|
| env0 | Windows Server 2022 (Victim) + Kali Linux (Attacker) |
Chains are not guaranteed to run to completion, some steps fail at execution time even when the plan is well-formed. We've catalogued the recurring causes in FAILURE_MODES.md. We recommend users read it to better understand Aurora.
We welcome community review of the attack chains and the underlying AALM predicates (preconditions / effects in each attack_chain.yml). If you spot inaccuracies, edge cases, or think a chain could be modeled better, please open a GitHub issue or PR. Discussion is what will make this dataset useful over time.
If you use Aurora or the chains in this repo, please cite:
@article{wang2024sands,
title={{From Sands to Mansions: Towards Automated Cyberattack Emulation with Classical Planning and Large Language Models}},
author={Wang, Lingzhi and Li, Zhenyuan and Jiang, Yi and Wang, Zhengkai and Guo, Zonghan and Wang, Jiahui and Wei, Yangyang and Shen, Xiangmin and Ruan, Wei and Chen, Yan},
journal={arXiv preprint arXiv:2407.16928},
year={2024},
note={Accepted to ACNS 2026}
}(The ACNS 2026 proceedings citation will be added once the conference publishes.)
Distributed under the Apache License 2.0. See LICENSE for details.
The attack chains in this repository are for education, research, and defensive testing purposes only. The authors do not condone unauthorized use. Use at your own risk.
- Lingzhi Wang — lingzhiwang2025@u.northwestern.edu
- Yi Jiang — jiangyi99@zju.edu.cn
- Zhengkai Wang — wangzhengkai@zju.edu.cn