language:

• en
  license: apache-2.0
  tags:
• cybersecurity
• red-team
• penetration-testing
• security
• offensive-security
• distillation
  size_categories:
• 1K<n<10K
  task_categories:
• text-generation
  pretty_name: Red Team Dataset

Red Team Dataset

A structured cybersecurity dataset where each example is a complete reasoning trajectory — a realistic sequence of steps and explanations that an AI assistant would produce during an authorized red team engagement.

Overview

This dataset contains 8,889 red team examples across 26 offensive security topics. Unlike traditional Q&A datasets, each row is a complete reasoning trajectory where an AI assistant plans an attack, explains methodology, reasons about findings, and chains multiple steps together to achieve an objective.

Each prompt embeds a scenario description, so the model knows the context, environment, and goal. The completions contain full natural language trajectories with structured reasoning and step-by-step methodology.

Dataset Statistics

What Makes This Different

Traditional security datasets provide Q&A pairs — a question and a text answer.

This dataset provides complete reasoning workflows:

Scenario: "Map the attack surface of an internal network from a compromised host"

↓
Assistant thinks: "First I need to enumerate the system and network"
↓
Explains enumeration steps and commands
↓
Assistant thinks: "Now I will identify privilege escalation paths"
↓
Explains escalation techniques and checks
↓
Assistant thinks: "Next I will pivot to other systems"
↓
Explains lateral movement methodology
↓
...continues through full attack chain

Topic Coverage

How It Was Made

Prompt Design

Each of the examples contains two parts:

1. Scenario: A realistic red team scenario with target IPs, domains, industries, OS versions, and objectives. Variable substitution ensures uniqueness across prompts.

2. Instruction Context: A structured prompt guiding the model to produce step-by-step reasoning and methodology.

Generation

Prompts were processed through a data generation pipeline using a high-capability language model at scale with high reasoning effort. The system prompt instructs the model to output complete reasoning trajectories with:

• Structured step-by-step explanations
• Realistic attack methodology
• Logical sequencing of actions
• Consistent technical details across steps

Self-Verification

The verification pipeline sends each generated trajectory back to the model in a fresh, independent chat for quality review:

1. Verification checks: Logical inconsistencies, shallow reasoning, hallucinated details, incomplete methodology, and structural issues.

2. Persistent pass set: Rows that pass verification are locked and never re-checked. Only failed rows get regenerated and re-verified.

3. ERROR vs FAIL separation: Generation errors retry automatically. Content failures trigger regeneration.

4. Result: The final dataset consists of high-quality verified responses after multiple rounds of refinement.

Quality Requirements

The system prompt and verification both enforce:

• Logical attack ordering
• Consistent reasoning across steps
• Realistic methodology
• Sufficient depth (multi-step explanations)
• Clear structure and completeness

Data Format

Each row is a JSONL object. The messages array contains the system prompt, user scenario, and assistant reasoning trajectory:

{
  "messages": [
    {
      "role": "system",
      "content": "You are an expert red-team AI assistant conducting authorized security assessments..."
    },
    {
      "role": "user",
      "content": "SCENARIO: During an authorized penetration test..."
    },
    {
      "role": "assistant",
      "content": "First, I will begin by enumerating the system..."
    }
  ],
  "topic": "active_directory_attacks"
}

Usage

from datasets import load_dataset

dataset = load_dataset("sh111111111111111/red_team")

# Filter by topic
cloud = dataset["train"].filter(lambda x: x.get("topic") == "cloud_penetration_testing")

# Extract reasoning responses
for row in dataset["train"]:
    assistant = row["messages"][2]["content"]
    print(len(assistant.split()))

Intended Use

• Fine-tuning security models: Train AI assistants to produce structured red team reasoning
• Instruction-following training: Improve reasoning depth in security tasks
• Red team automation research: Model multi-step attack reasoning
• Security education: Demonstrate realistic attack chains with detailed explanations

All scenarios are framed within authorized penetration tests, red team engagements, or security assessments.

Limitations

• Self-verification bias: The same model generated and verified responses
• No real execution: Outputs are synthetic and not from live systems

Technical Details

Citation

@dataset{red_team_2026,
  title={Red Team Dataset},
  year={2026},
  url={https://huggingface.co/datasets/sh111111111111111/red_team},
  note={8,889 verified red team reasoning trajectories across 26 topics}
}

License

Apache 2.0