language:
- en
license: mit
task_categories: - text-generation
- text-classification
tags: - cyber-security
- red-teaming
- chain-of-thought
- agent-reasoning
- synthetic
dataset_info:
features:- name: instruction
dtype: string - name: response
dtype: string - name: hypothesis
dtype: string - name: confirmed
dtype: bool - name: severity
dtype: string
splits: - name: train
num_bytes: 73400320
num_examples: 100000
download_size: 73400320
dataset_size: 73400320
configs:
- name: instruction
- config_name: default
data_files:- split: train
path: security_dataset.jsonl
- split: train
🛡️ Security Analyst CoT Dataset (100k)
A massive-scale, synthetically generated dataset designed to train AI Security Agents in offensive reasoning, vulnerability verification, and false positive reduction.
Dataset Summary
- Size: 100,000 Unique Samples
- Format: JSONL
- Focus: Chain-of-Thought (CoT) Reasoning for Web Security
- Logic: Observation -> Hypothesis -> Evidence -> Decision (O-H-E-D)
Features
Each sample simulates a complete cognitive process of a Senior Security Analyst:
- Instruction: The raw HTTP request + The server's response code/body.
- Reasoning: A `
` block analyzing the anomaly, evidence, and conclusion.
3. **Verdict**: Structured labels for `STATUS`, `SEVERITY`, and `NEXT_TEST`.
## Attack Categories
The dataset covers a wide spectrum of modern web threats:
- **Injection**: SQLi, XSS (Reflected/Stored), Command Injection, LDAPi.
- **Business Logic**: Price Manipulation, Mass Assignment, Race Conditions.
- **Protocol**: HTTP Request Smuggling, Host Header Injection.
- **Anomalies**: Zero-Day simulations (Unknown patterns) and Fuzzing noise.
- **Benign**: High-entropy legitimate traffic to train False Positive rejection.
## Sample Structure
```json
{
"instruction": "Analyze: GET /api/v1/user?id=1' OR 1=1 HTTP/1.1...",
"response": "\nObservation: ...\nHypothesis: ...\nEvidence: ...\nDecision: ...\n
3. **Verdict**: Structured labels for `STATUS`, `SEVERITY`, and `NEXT_TEST`.
## Attack Categories
The dataset covers a wide spectrum of modern web threats:
- **Injection**: SQLi, XSS (Reflected/Stored), Command Injection, LDAPi.
- **Business Logic**: Price Manipulation, Mass Assignment, Race Conditions.
- **Protocol**: HTTP Request Smuggling, Host Header Injection.
- **Anomalies**: Zero-Day simulations (Unknown patterns) and Fuzzing noise.
- **Benign**: High-entropy legitimate traffic to train False Positive rejection.
## Sample Structure
```json
{
"instruction": "Analyze: GET /api/v1/user?id=1' OR 1=1 HTTP/1.1...",
"response": "
\nSTATUS: VULNERABLE...",
"hypothesis": "SQL Injection",
"confirmed": true,
"severity": "Critical"
}