---
dataset_info:
  features:
    - name: messages
      list:
        - name: role
          dtype: string
        - name: content
          dtype: string
    - name: id
      dtype: string
license: apache-2.0
task_categories:
  - text-generation
tags:
  - cybersecurity
  - pentesting
  - active-directory
  - synthetic-data
  - openworlds
pretty_name: OpenWorlds AD Penetration Testing Trajectories
---

# OpenWorlds AD Penetration Testing Trajectories

Synthetic training trajectories for fine-tuning AI agents on Active Directory penetration testing.

## Dataset Description

Each trajectory is a complete attack path from **initial access** to **Domain Admin**, formatted as multi-turn chat conversations with:

- **System prompt** with target domain context
- **<think> reasoning** traces (expert thought process)
- **Tool calls** (nmap, ldapsearch, GetUserSPNs, hashcat, secretsdump, etc.)
- **Tool outputs** (realistic simulated responses)
- **Failure recovery** (typos, wrong creds, with correction steps)

## Attack Strategies Covered

| Strategy | Description |
|----------|-------------|
| Kerberoasting | Service accounts with SPNs -> crack TGS tickets |
| AS-REP Roasting | Users without pre-auth -> crack AS-REP hashes |
| ACL Abuse Chains | GenericAll -> WriteDACL -> ForceChangePassword -> DCSync |
| AD CS Abuse (ESC1) | Vulnerable cert templates -> impersonate Domain Admin |
| Credential Pivoting | Passwords in shares, GPPs, config files |

## Usage

```python
from datasets import load_dataset
ds = load_dataset("omkar6699/openworlds-ad-trajectories")
```

## Generation

Generated using [OpenWorlds](https://github.com/omkar-ukirde/Worlds-OpenSource-Clone):

```bash
openworlds manifest generate --hosts 10 --users 25 --seed 42
openworlds trajectory generate --failure-rate 0.15
```

## License

Apache 2.0