AD-GEN: Evidence-Preserving Generation of Validated ATT&CK-Aligned Narratives from Large-Scale Endpoint Telemetry

LLM-ready SOC narratives derived from large-scale Windows endpoint telemetry

---

Overview

Modern endpoint telemetry datasets contain rich behavioral evidence but are often difficult to use directly for large language model (LLM) reasoning. Raw Sysmon logs are fragmented across individual events, contain sensitive identifiers, suffer from process identifier reuse, and provide limited behavioral context for security analysis.

AD-GEN addresses this gap by transforming large-scale Windows Sysmon telemetry from the COMISET corpus into process-centric, privacy-preserving, compressed, and validated ATT&CK-aligned narrative records. The framework reconstructs process lifecycles, preserves behavioral evidence, normalizes temporal information, and generates analyst-style annotations suitable for LLM-based cybersecurity applications.

AD-GEN is designed for:

• LLM-based SOC automation
• ATT&CK-aware instruction tuning
• Threat hunting assistant development
• Endpoint behavior reasoning
• Security narrative generation

The resulting dataset enables security-focused language models to reason over structured behavioral narratives rather than isolated raw events, while maintaining alignment with the MITRE ATT&CK framework through automated validation.

Dataset Scale

Metric	LAB	REAL	Total
Raw Sysmon Events	49,914,325	202,304,790	252,219,115
Post-Squash Events	21,360,985	31,571,618	52,932,603
Step 2 Narratives	49,745	185,978	235,723
Step 3 LLM Outputs	50,671	190,109	240,780
Step 4 Validated Outputs	50,622	190,085	240,707

---

Compression Statistics

Environment	Raw Events	Post-Squash Events	Compression Ratio	Event Reduction
LAB	49,914,325	21,360,985	2.34×	57.20%
REAL	202,304,790	31,571,618	6.41×	84.39%
Overall	252,219,115	52,932,603	4.76×	79.01%

---

Risk Distribution

Risk Level	Count	Percentage
Low	234,046	97.26%
Medium	3,131	1.30%
High	2,607	1.08%
Critical	847	0.35%

---

Top MITRE ATT&CK Tactics

Tactic ID	Tactic Name	Frequency
TA0004	Privilege Escalation	3,157
TA0005	Defense Evasion	2,511
TA0003	Persistence	2,451
TA0002	Execution	1,763
TA0007	Discovery	727
TA0006	Credential Access	603

---

Output Format

Each record is stored as JSONL.

{
  "sample_id": "ADGEN_0000001",
  "environment": "LAB",
  "source_dataset": "COMISET",
  "narrative": "...",
  "sysmon_hints": ["T1055"],
  "label": {
    "risk_level": "Medium",
    "mitre_tactics": ["TA0005"],
    "mitre_techniques": ["T1055"],
    "recommended_actions": [
      {
        "tool_name": "get_file_metadata",
        "parameters": {}
      }
    ],
    "summary": "Process access behavior consistent with process injection.",
    "analyst_rationale": "Evidence-based analyst reasoning.",
    "verdict": "suspicious",
    "label_source": "llm_validated"
  }
}

---

Supported SOC Actions

check_threat_intel
get_file_metadata
query_registry
get_network_flow
terminate_process
isolate_host
no_action

---

Label Quality

Metric	LAB	REAL
Parse Success	100.00%	100.00%
Schema Validity	99.93%	99.98%
Verdict Consistency	95.98%	98.64%
Unknown Tactics After Validation	0.032%	0.007%
Unknown Techniques After Validation	0.041%	0.013%
Invalid Actions	0	0

---

Cross-Model Audit

Metric	Value
Audit Samples	300
Auditor Models	3
GPT-5.5 Composite Score	0.748
Claude Opus 4.8 Composite Score	0.748
Minimum Evidence Support Score	>0.72
Minimum ATT&CK Alignment Score	>0.72

Three independent frontier language models (GPT-5.5, Claude Opus 4.8, and Gemini 3.5 Flash) were used to evaluate stratified samples from both LAB and REAL environments. Independent audits converged to nearly identical quality assessments, supporting the reliability and consistency of the validated ATT&CK-aligned labels.

Repository Structure

AD-GEN
├── README.md
├── LAB
│   └── NEW_LAB.jsonl
├── REAL
│   └── NEW_REAL.jsonl
├── docs
│   └── pipeline.png
└── Conversion
    ├── Conversion.py
    └── react_soc_prompt.txt

---

Installation

git clone https://github.com/namhop88/AD-GEN.git
cd AD-GEN

---

Files

File	Description
LAB/NEW_LAB.jsonl	AD-GEN processed records derived from the COMISET laboratory environment
REAL/NEW_REAL.jsonl	AD-GEN processed records derived from the COMISET real university network environment
docs/pipeline.png	Overview of the AD-GEN transformation pipeline
Conversion/Conversion.py	Utility for converting records to the release format
Conversion/react_soc_prompt.txt	ReAct-style SOC labeling prompt used during generation

---

AD-GEN is derived from the COMISET Windows endpoint telemetry corpus; it does not claim to be the original raw telemetry source.

Important Note

AD-GEN labels are validated synthetic analyst labels, not human-adjudicated forensic ground truth.

The dataset is intended for research in instruction tuning, weak supervision, SOC assistant development, ATT&CK-aware reasoning, and endpoint narrative modeling. Additional expert review is recommended before operational use.

---

Citation

@article{nam2026adgen,
  title   = {AD-GEN: Evidence-Preserving Generation of Validated ATT\&CK-Aligned Narratives from Large-Scale Endpoint Telemetry},
  author  = {Dinh Phuong Nam and Nguyen Tan Cam},
  year    = {2026},
  journal = {Preprint}
}

---

License

Component	License
Source Code	MIT License
Dataset	CC BY-NC 4.0

---

Disclaimer

AD-GEN is released for academic and defensive cybersecurity research only. It should not be used as the sole basis for operational security decisions without expert validation.

---

Contact

Dinh Phuong Nam
University of Information Technology (UIT), VNU-HCM

HUTECH University