language:
- en
license: mit
task_categories: - text-classification
- token-classification
tags: - security
- cybersecurity
- mitre-attack
- microsoft-sentinel
- kql
- adversarial
- red-team
- detection-engineering
- synthetic
size_categories: - 1K<n<10K
DUEL Adversarial Security Dataset
Dataset Description
This dataset contains synthetic adversarial telemetry generated by the DUEL framework
(Dual Unified Evasion Loop) — an adversarial LLM security research framework where an
Attacker agent and a Defender agent battle across MITRE ATT&CK and OWASP LLM Top 10
techniques against real Microsoft Sentinel schemas.
Every record is a single synthetic log entry from one round of the adversarial loop,
labelled as evaded (the Defender's KQL rule missed it) or detected (caught by the rule).
The accompanying kql_rule field contains the Defender's detection rule for that round.
This dataset is intended for:
- Training and evaluating ML-based intrusion detection models
- Benchmarking KQL detection rule quality
- Studying adversarial mutation strategies in cloud security
- Red/blue team simulation and detection engineering research
How It Was Generated
Attacker (llama3.1:8b) Defender (mistral:7b)
| generates synthetic telemetry | writes KQL detection rule
| mutates each round based on | hardens rule each round based
what got detected on what evaded
<-> Detection Engine (pandas KQL executor) <->
scores every round -> labels: evaded / detected
Each battle runs multiple rounds per technique. The Attacker carries persistent memory
across battles — evasion patterns accumulate and feed future mutation strategies.
Dataset Statistics
| Split | Records |
|---|---|
| Train (70%) | 726 |
| Validation (15%) | 155 |
| Test (15%) | 157 |
| Total | 1038 |
Techniques covered: 38
LLM01, LLM02, LLM03, LLM04, LLM05, LLM06, LLM07, LLM08, LLM09, LLM10, T1018, T1040, T1069.003, T1078.001, T1078.004, T1087.004, T1098.001, T1110.003, T1114.002, T1133, T1136.003, T1190, T1199, T1485, T1486, T1526, T1528, T1530, T1537, T1550.001, T1556.006, T1562.001, T1566.001, T1566.002, T1567.002, T1606.002, T1621, T1648
Label Distribution
| Label | Count |
|---|---|
evaded |
620 |
detected |
418 |
Average evasion rate: 60.9%
Evasion Rate Distribution
| Evasion Rate Bin | Records |
|---|---|
| 0-25% | 300 |
| 25-50% | 76 |
| 50-75% | 144 |
| 75-100% | 518 |
Field Descriptions
| Field | Type | Description |
|---|---|---|
id |
string | UUID for this record |
technique_id |
string | MITRE ATT&CK or OWASP LLM technique ID (e.g., T1078.004, LLM01) |
technique_name |
string | Human-readable technique name |
tactic |
string | ATT&CK tactic (e.g., Initial Access, Persistence) |
round |
int | Battle round number (1-indexed) |
log |
object | Synthetic telemetry entry (Microsoft Sentinel schema) |
label |
string | evaded or detected — outcome against the Defender's KQL rule |
kql_rule |
string | KQL detection rule used by the Defender in this round |
attacker_strategy |
string | Description of the Attacker's mutation strategy for this round |
defender_reasoning |
string | Description of the Defender's hardening approach for this round |
evasion_rate |
float | Fraction of logs that evaded detection this round (0.0-1.0) |
mutation_round |
int | Same as round — the mutation iteration number |
model_attacker |
string | Ollama model used for the Attacker (e.g., llama3.1:8b) |
model_defender |
string | Ollama model used for the Defender (e.g., mistral:7b) |
Sentinel Table Schemas
The log field conforms to real Microsoft Sentinel table schemas:
SigninLogs— Azure AD sign-in eventsAuditLogs— Azure AD audit eventsAzureActivity— Azure resource management operationsOfficeActivity— Microsoft 365 activity events- Custom OWASP LLM schemas for AI/LLM attack simulation
Citation
@software{duel_framework_2026,
author = {Daniel Gomes},
title = {DUEL: Dual Unified Evasion Loop -- Adversarial LLM Security Research Framework},
year = {2026},
url = {https://github.com/0xDanielSec/duel-framework},
note = {Adversarial telemetry dataset generated by LLM agents battling across
MITRE ATT&CK and OWASP LLM Top 10 techniques}
}
Framework
Generated by DUEL — an open-source
adversarial LLM security research framework. Zero external API calls — fully local via Ollama.
License
MIT — see LICENSE.
All telemetry is entirely synthetic. No real Azure tenants, Microsoft 365 environments,
or live Sentinel workspaces are involved.