license: mit
language:
- en
pretty_name: Offensive Security Training Data Samples
tags: - cybersecurity
- red-team
- penetration-testing
- offensive-security
- vulnerability-research
- llm-training
task_categories: - text-generation
size_categories: - n<10K
Offensive Security — Training Data Samples
Preview samples from the Lateos Red Team training pipeline. Contains a deterministic 3% sample (seed 42) of the full corpus, drawn from the exact production datasets used to train red-team and offensive-security models.
These samples let you evaluate schema, quality, and provenance before licensing the full datasets. All data is dual-use security research: attack playbooks, vulnerability analysis, tool usage guides, and decision frameworks for authorized penetration testing.
Why this dataset exists
Generic LLMs lack structured knowledge of authorized offensive security operations — exploit techniques, red team methodologies, vulnerability classification, and operational decision-making are domains where public web data is fragmented and often inaccurate. This corpus grounds a model in:
- Attack taxonomies: MITRE ATT&CK technique descriptions, kill chain phases, and adversarial TTPs
- Vulnerability databases: CVE records, exploit patterns, severity classifications, and CWE mappings
- Tool reference materials: Usage guides for industry-standard tools (Metasploit, Nmap, Burp Suite, Cobalt Strike)
- Operational frameworks: Rules of engagement, target analysis strategies, safety checks, and autonomous decision trees
- Instruction-tuning data: Structured Q&A pairs covering reconnaissance, exploitation, post-exploitation, and reporting
Configurations
| Config | File | Records (3%) | Schema |
|---|---|---|---|
default |
offensive_security_sample.parquet |
269 | instruction, input, output, provenance |
What it trains
- Structured vulnerability analysis — Each record pairs a CVE or vulnerability with attack vector, severity, classification, and defensive actions. Trains the model to turn raw advisories into attributed, actionable findings.
- Exploit technique reasoning — How specific vulnerabilities map to exploitation steps, required conditions, and expected outcomes.
- Tool orchestration — When to invoke specific tools, how to structure commands, and how to interpret outputs.
- Red team methodology — Phased approach from reconnaissance through reporting, including decision points and safety boundaries.
- DeFi/Smart Contract security — Vulnerability patterns specific to blockchain and decentralized finance applications.
Provenance
- All samples are deterministic subsets (seed 42) of the production, audit-clean v2 corpora
- Full corpora ship with per-record provenance, record-id remap files, stage hashes, and reproducibility scripts
- Generator: Grok Build 0.1 — all records reviewed and enriched via automated quality pipeline
- Quality grade: gold — zero echo records, zero parse errors, all silver-class defects repaired
Access
The full datasets are available under license terms. Contact [email protected] or visit https://lateos.ai/categories/datasets/.
Citation
@misc{lateos_offensive_security,
title = {Offensive Security Training Data Samples},
author = {Lateos Inc.},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/Lateos/offensive_security_dataset}
}