README.md

---
license: apache-2.0
task_categories:
  - text-classification
  - graph-ml
language:
  - en
tags:
  - cybersecurity
  - intrusion-detection
  - provenance-graphs
  - MITRE-ATT&CK
  - SOAR
  - security-operations
  - IDS
  - network-security
  - threat-detection
  - labeled-dataset
  - lead-rules
size_categories:
  - 100M<n<1B
dataset_info:
  - config_name: signals
    splits:
      - name: train
        num_examples: 114234041
configs:
  - config_name: signals
    data_files:
      - split: train
        path: signals/*.parquet
  - config_name: graph_nodes
    data_files:
      - split: train
        path: graph/nodes.jsonl
  - config_name: graph_edges
    data_files:
      - split: train
        path: graph/edges.jsonl
  - config_name: incidents
    data_files:
      - split: train
        path: graph/incidents.jsonl
---

# WitFoo Precinct6 Cybersecurity Dataset (large)

## Overview

A large-scale, labeled cybersecurity dataset derived from production Security Operations Center (SOC) data processed by [WitFoo Precinct](https://www.witfoo.com/) version 6.x. This dataset contains **114,234,041 sanitized security events** (signal logs) across 5 organizations and **12,361 incident provenance graphs** (47,632 nodes, 32,086,552 edges).

**Available in two sizes:**
- [`witfoo/precinct6-cybersecurity`](https://huggingface.co/datasets/witfoo/precinct6-cybersecurity) — 2.1M signals (smaller, faster to load; same incidents and graph methodology)
- [`witfoo/precinct6-cybersecurity-100m`](https://huggingface.co/datasets/witfoo/precinct6-cybersecurity-100m) — **114M signals (this dataset)**

**Generate your own:** WitFoo Precinct 6.x customers can create datasets from their own data using the open-source pipeline: [`witfoo/dataset-from-precinct6`](https://github.com/witfoo/dataset-from-precinct6)

This dataset is designed to support research in:
- **Provenance graph-based intrusion detection** (KnowHow, NodLink, and similar systems)
- **AI-driven cyber defense simulation** (CybORG and MARL-based defense policy training)
- **Security alert classification** (malicious vs. suspicious vs. benign event labeling)
- **Attack lifecycle analysis** using MITRE ATT&CK framework mappings
- **Detection rule evaluation** using WitFoo's 261 lead detection rules

## Quick Start

```python
from datasets import load_dataset

# Load flat signal logs (114M rows across 58 Parquet shards)
signals = load_dataset("witfoo/precinct6-cybersecurity-100m", "signals", split="train")

# Find analyst-confirmed malicious events
malicious_confirmed = signals.filter(
    lambda x: x["label_binary"] == "malicious" and x["disposition"] == "Disrupted"
)

# Find suspicious events that matched detection rules but are not in confirmed incidents
suspicious = signals.filter(lambda x: x["label_binary"] == "suspicious")

# Load provenance graph (47.6k nodes, 32M edges)
nodes = load_dataset("witfoo/precinct6-cybersecurity-100m", "graph_nodes", split="train")
edges = load_dataset("witfoo/precinct6-cybersecurity-100m", "graph_edges", split="train")

# Load full incident graphs (12,361 incidents with embedded artifacts, leads, frameworks)
incidents = load_dataset("witfoo/precinct6-cybersecurity-100m", "incidents", split="train")
```

## Label Distribution

| Label | Count | Percentage |
|-------|-------|------------|
| `benign` | 113,543,372 | 99.40% |
| `suspicious` | 616,605 | 0.54% |
| `malicious` | 74,064 | 0.06% |

Disposition breakdown (the raw Precinct status, exposed for ground-truth stratification — see [Ground Truth](#ground-truth-and-disposition)):

| Disposition | Count | Meaning |
|-------------|-------|---------|
| `Unprocessed` | 114,192,640 | No analyst review (default state for benign/suspicious + un-reviewed malicious) |
| `Disrupted` | 41,391 | SOC analyst confirmed and intervened |
| `Dismissed` | 10 | SOC analyst dismissed |

## Signal Columns

| Column | Type | Description |
|--------|------|-------------|
| `timestamp` | float | Unix epoch timestamp |
| `message_type` | string | Event classification (e.g., `firewall_action`, `account_logon`, `AssumeRole`) |
| `stream_name` | string | Source product/data stream |
| `pipeline` | string | Ingestion pipeline |
| `src_ip` | string | Source IP (sanitized) |
| `dst_ip` | string | Destination IP (sanitized) |
| `src_port` | string | Source port |
| `dst_port` | string | Destination port |
| `protocol` | string | Network protocol (6=TCP, 17=UDP, 1=ICMP) |
| `src_host` | string | Source hostname (sanitized) |
| `dst_host` | string | Destination hostname (sanitized) |
| `username` | string | Associated username (sanitized) |
| `action` | string | Event action (block, permit, logon, logoff) |
| `severity` | string | Severity level |
| `vendor_code` | string | Vendor-specific event code |
| `message_sanitized` | string | Full sanitized raw log message |
| `label_binary` | string | `malicious`, `suspicious`, or `benign` |
| `label_confidence` | float | Confidence score (0.0–1.0). See [Scoring](#scoring). |
| `attack_techniques` | string | JSON array of MITRE ATT&CK technique IDs (e.g., `["T1041","T1567"]`) |
| `attack_tactics` | string | JSON array of MITRE ATT&CK tactic IDs (`TA0001`-style) |
| `defense_techniques` | string | JSON array of MITRE D3FEND defense technique IDs |
| `suspicion_score` | float | WitFoo suspicion score (0.0–1.0). See [Scoring](#scoring). |
| `mo_name` | string | Modus operandi (e.g., `Data Theft`) |
| `lifecycle_stage` | string | Kill chain stage (e.g., `initial-compromise`, `complete-mission`) |
| `disposition` | string | Raw Precinct status (`Disrupted`, `Investigating`, `Resolved`, `Dismissed`, `False Positive`, `Unprocessed`) |
| `disposition_category` | string | Bucketed disposition (`confirmed-malicious`, `false-positive`, `dismissed`, `automated`) |
| `is_false_positive` | bool | True if SOC analyst marked the incident as a false positive |
| `status_name` | string | Same as `disposition` (raw Precinct status) |
| `incident_ids` | string | JSON array of incident UUIDs (empty for benign/suspicious) |
| `matched_rules` | string | JSON array of matched WitFoo lead rule descriptions |
| `set_roles` | string | JSON array of classification roles (e.g., `Exploiting Host`, `C2 Server`) |
| `product_name` | string | Security product name (e.g., `ASA Firewall`, `Falcon`) |
| `vendor_name` | string | Product vendor (e.g., `Cisco`, `Crowdstrike`) |

## Source Products

The dataset contains events from **158 security products** across **70+ vendors**. Complete catalog in `reference/lead_rules_catalog.json`.

| Category | Products |
|----------|----------|
| **Firewalls** | Cisco ASA, Palo Alto PAN NGFW, Fortinet FortiGate, Checkpoint, Meraki, SonicWall, pfSense, Barracuda, Juniper SRX |
| **Endpoint Protection** | CrowdStrike Falcon, Symantec SEP, Carbon Black, Cylance, SentinelOne, Deep Instinct, Sophos, McAfee, ESET |
| **Network Detection** | Cisco Stealthwatch, Cisco Firepower, Suricata IDS, TippingPoint IPS, Vectra Cognito |
| **Identity & Access** | Microsoft Windows AD, Cisco ISE, Centrify, CyberArk, Duo, Okta, Beyond Trust |
| **Cloud Security** | AWS CloudTrail, AWS VPC Flow Logs, AWS GuardDuty, Azure Security, Zscaler, Netskope, Cisco Umbrella |
| **Email Security** | ProofPoint, Mimecast, FireEye EX, Barracuda ESS, Cisco IronPort, Checkpoint Harmony |
| **Threat Intelligence** | FireEye NX/HX/AX/CMS, Trend Micro, QRadar, Microsoft ATA, Cortex XDR |
| **Infrastructure** | VMware vCenter/NSX, Elastic Filebeat, Linux (sshd, PAM, systemd, auditd), Apache, HAProxy |

## Labeling Methodology

**Three-tier labels** derived from two sources:

- **`malicious`** (74,064): Events embedded as leads inside confirmed incidents. Extracted directly from incident lead objects with suspicion scores, modus operandi, and MITRE mappings.
- **`suspicious`** (616,605): Events matching WitFoo's 261 lead detection rules but not present in confirmed incidents.
- **`benign`** (113,543,372): Events not matching any detection rules and not in any incident.

### Ground Truth and Disposition

**All labels in this dataset are derived from WitFoo Precinct's automated incident correlation engine — there is no independent, analyst-verified ground truth.** Researchers should treat Precinct's analysis as a strong but imperfect oracle. The `disposition` column lets you assess label quality on a per-record basis:

| `disposition` | Meaning | Confidence in label |
|---------------|---------|---------------------|
| `Disrupted` | SOC analyst confirmed the incident and intervened | High — human-confirmed malicious |
| `Investigating` | SOC analyst is actively investigating | Medium — analyst engaged |
| `Resolved` | SOC analyst confirmed and resolved | High — human-confirmed malicious |
| `Dismissed` | SOC analyst dismissed the incident | Negative — analyst rejected |
| `False Positive` | SOC analyst confirmed false positive | Negative — analyst rejected |
| `Unprocessed` | Automated detection, no human review | Lower — Precinct-confidence only |

The `disposition_category` column buckets these into four values for easier filtering: `confirmed-malicious`, `false-positive`, `dismissed`, `automated`. For experiments where ground-truth quality matters, restrict to records where `disposition` ∈ {`Disrupted`, `Resolved`} to compare against analyst-confirmed labels. **This dataset has 41,391 records with `Disrupted` status, 10 with `Dismissed`, and the remainder with `Unprocessed` (no analyst review).**

For benign and suspicious records, `disposition` is `Unprocessed` (no incident association). For malicious records, `disposition` reflects the parent incident's status at extraction time.

### Scoring

The dataset exposes two related score fields:

- **`suspicion_score`** (float, 0.0–1.0) — Precinct's proprietary suspicion score from the parent incident. Populated for malicious records; zero for benign and suspicious. (Previously this column was always `0` — that was a defaulting bug, now fixed.)

- **`label_confidence`** (float, 0.0–1.0) — Confidence in the assigned `label_binary` tier. Computed deterministically from corroborating signal:

  | Label | Formula |
  |-------|---------|
  | `malicious` | `max(0.6, suspicion_score)` clamped to 0.95; lowered to 0.3 if `is_false_positive` |
  | `suspicious` | `0.4 + 0.1 × n_matched_rules + 0.05 × n_set_roles`, clamped to [0.5, 0.85] |
  | `benign` | `0.5` (no positive evidence either way) |

  Note: `label_confidence` is **not** the probability the activity is malicious — it indicates how much corroborating evidence supports the assigned tier. Source: [`src/precinct6_dataset/label.py`](https://github.com/witfoo/dataset-from-precinct6/blob/main/src/precinct6_dataset/label.py).

### MITRE ATT&CK Mappings

Attack technique and tactic labels are derived from three sources, with deduplication:

1. **WitFoo set role names** attached to the incident (e.g., `C2 Server` → `TA0011` Command and Control, `T1071` Application Layer Protocol)
2. **Modus operandi** name on the incident (e.g., `Ransomware` → `TA0001`, `TA0002`, `TA0040`; `T1486` Data Encrypted for Impact)
3. **Per-product framework data** embedded in `incident.nodes.products.frameworks` (when present)

Tactic IDs use the standard MITRE ATT&CK Enterprise format (`TA0001` through `TA0043`). Technique IDs are top-level techniques representing the most likely category for a given role. Researchers wanting precise per-event technique attribution should treat these as priors. Full mapping tables in [`src/precinct6_dataset/mitre_mapping.py`](https://github.com/witfoo/dataset-from-precinct6/blob/main/src/precinct6_dataset/mitre_mapping.py).

**Per-edge/per-node attribution in graph output:** `attack_tactics`, `attack_techniques`, `set_roles`, `lifecycle_stage`, `label_binary`, `label_confidence`, `suspicion_score`, `disposition` are attached at the **edge** level in `edges.jsonl` and in per-incident GraphML files. Nodes in `incidents.jsonl` carry their own `sets` and `products` dicts with per-entity information.

## Graph Data

| Component | Count |
|-----------|-------|
| Nodes (artifact-derived hosts + credentials) | 47,632 |
| Edges (artifact `EVENT`/`NETWORK_FLOW` + incident `INCIDENT_LINK`) | 32,086,552 |
| Incidents | 12,361 |

Per-edge labels in `edges.jsonl` include `attack_techniques`, `attack_tactics`, `set_roles`, `lifecycle_stage`, `disposition`, `mo_name`, `suspicion_score`, `incident_id` (for INCIDENT_LINK edges).

## Attack Reports

`graph/attack_reports.jsonl` contains one natural-language threat-hunting report per incident (**12,361 reports**). Each report is deterministically composed from structured incident metadata (modus operandi, set roles, lead descriptions, MITRE mappings, timestamps) and explicitly states that it **reflects Precinct's automated correlation engine output, not an independent threat-hunting investigation**.

Each record contains: `incident_id`, `report_text`, `mo_name`, `suspicion_score`, `disposition`, `attack_techniques`, `attack_tactics`, `lead_count`, `set_role_names`, `matched_rules`, `products_observed`, `lifecycle_stage`, and timing fields.

Researchers can audit exactly how each sentence is derived by reading [`src/precinct6_dataset/attack_reports.py`](https://github.com/witfoo/dataset-from-precinct6/blob/main/src/precinct6_dataset/attack_reports.py).

## Additional Files

- **`signals/signals-NNNNN.parquet`** — 58 Parquet shards (~270 MB each), 2M rows per shard except final shard. Total ~15.7 GB.
- **`signals/metadata.json`** — Per-shard summary, label/disposition distributions, top message types and streams.
- **`graph/nodes.jsonl`** — 47,632 graph nodes (hosts + credentials derived from artifact src/dst/username).
- **`graph/edges.jsonl`** — 32,086,552 graph edges, each with `labels` dict carrying MITRE, disposition, set_roles, etc.
- **`graph/incidents.jsonl`** — Full incident records (12,361 lines, ~638 MB) with embedded `nodes`, `edges`, `leads`, and framework mappings.
- **`graph/incidents_graphml/{0-f}/{incident_id}.graphml`** — **12,361 per-incident GraphML files**, sharded into 16 subdirectories by first hex char of the incident UUID (HuggingFace caps directories at 10,000 files). Each file is small (KB-MB) and loadable in Gephi, NetworkX, igraph, or DGL. Ideal for graph-based research where loading the entire dataset is impractical.
- **`graph/attack_reports.jsonl`** — Natural-language threat-hunting reports (12,361 reports, ~21 MB). See [Attack Reports](#attack-reports).
- **`reference/lead_rules_catalog.json`** — Complete catalog of 261 WitFoo lead detection rules, 158 security products, 106 classification sets, and 216 stream-to-product mappings.

**Note on the artifact-level graph:** With 32M edges, the monolithic `graph.graphml` is intentionally **not** shipped at this scale — per-incident GraphMLs and the streaming `edges.jsonl` are the recommended entry points. Methodology and content fields are otherwise identical to the 2M version.

## Sanitization

All customer-identifying information has been removed through a comprehensive 4-layer sanitization pipeline. The pipeline is [open source](https://github.com/witfoo/dataset-from-precinct6) under the Apache 2.0 license.

1. **Structured field sanitization + Aho-Corasick multi-pattern sweep** — Known fields are replaced with deterministic tokens (IPs → [RFC 5737](https://datatracker.ietf.org/doc/html/rfc5737) ranges, hostnames → `HOST-NNNN`, etc.). Every record is then swept using an Aho-Corasick automaton built from 97,000+ PII registry entries.
2. **Format-specific log message parsing** — Eight specialized parsers handle Cisco ASA syslog, Windows Security Event XML, WinLogBeat JSON, AWS CloudTrail, Palo Alto Networks, VMware vCenter, DNS logs, and a generic fallback.
3. **Machine learning residual detection** — [Microsoft Presidio](https://microsoft.github.io/presidio/) (spaCy NLP) and [BERT NER](https://huggingface.co/dslim/bert-base-NER) scan for residual PII. New discoveries trigger full re-sanitization.
4. **Large language model contextual review** — [Claude](https://www.anthropic.com/claude) reviews stratified samples for subtle PII. Findings trigger re-sanitization.

The four layers run in cycles. PII discovered by ML/AI in one cycle is caught automatically by Layer 1 in all subsequent cycles. **Final PII registry: 97,046 unique mappings across 13 categories** (IPs, hostnames, usernames, orgs, credentials, SIDs, emails, ARNs, etc.). All replacements are consistent — the same original value always maps to the same token, preserving graph topology.

## Research Context

This dataset was produced in collaboration with the University of Canterbury (New Zealand) Computer Science and Software Engineering department for two research projects:
- **AI Cyber-Security Battle Simulator** — Improving CybORG with realistic IDS observations, graph-based defense policies, and AI-driven attacker modeling
- **Intrusion Detection based on Provenance Graphs** — Evaluating reproducibility and generalizability of KnowHow and NodLink detection methods

## Limitations

- **Label imbalance**: 99.40% benign reflects production SOC reality. Sampling strategies needed for balanced training.
- **Temporal scope**: July–August 2024.
- **Ground truth**: All labels derive from Precinct's automated correlation. Use the `disposition` column to stratify by analyst review level.
- **Shared incidents**: The same 12,361 incidents appear in both the 2M and this dataset (incidents are stored separately from signal data; only the signal sample size differs).
- **Sanitization trade-offs**: Some log message detail is reduced by PII replacement, particularly in free-text fields.

## Citation

```bibtex
@dataset{witfoo_precinct6_100m_2026,
  title={WitFoo Precinct6 Cybersecurity Dataset (large)},
  author={WitFoo, Inc.},
  year={2026},
  url={https://huggingface.co/datasets/witfoo/precinct6-cybersecurity-100m},
  license={Apache-2.0}
}
```

## License

[Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0)