license: mit
task_categories:
- text-classification
- question-answering
language: - en
tags: - security
- ai-safety
- incidents
- mitre-atlas
- vulnerabilities
- responsible-ai
size_categories: - 1K<n<10K
configs: - config_name: default
data_files:- split: incidents
path: data/incidents-* - split: incidents_atlas_mapped
path: data/incidents_atlas_mapped-* - split: jailbreaks
path: data/jailbreaks-*
dataset_info:
features: - name: id
dtype: string - name: source
dtype: string - name: source_id
dtype: string - name: title
dtype: string - name: description
dtype: string - name: incident_date
dtype: timestamp[s] - name: ingested_at
dtype: string - name: vendor
dtype: string - name: system
dtype: string - name: atlas_technique
dtype: string - name: atlas_tactic
dtype: string - name: severity
dtype: string - name: tags
list: string - name: url
dtype: string - name: raw
dtype: json
splits: - name: incidents
num_bytes: 10894973
num_examples: 4461 - name: incidents_atlas_mapped
num_bytes: 7411589
num_examples: 2725 - name: jailbreaks
num_bytes: 46469
num_examples: 15
download_size: 6689679
dataset_size: 18353031
- split: incidents
ai-sentinel-feed
Unified, normalized AI incident records from public sources, mapped where possible to the MITRE ATLAS adversarial ML taxonomy.
Live API: see the project README on GitHub (tk-tobi/ai-sentinel-feed).
Dataset splits
| Split | Description |
|---|---|
incidents |
All normalized records |
incidents_atlas_mapped |
Records with a mapped ATLAS technique (not unmapped) |
jailbreaks |
Records tagged heuristically for jailbreak / prompt-injection themes |
Record counts at last sync (2026-08-01T06:03:20Z):
{
"incidents": 4461,
"incidents_atlas_mapped": 2725,
"jailbreaks": 15
}
Total incidents in incidents: 4461.
Schema
Each row is a JSON object with:
| Field | Type | Description |
|---|---|---|
id |
string | Stable SHA-256 id (source + source_id) |
source |
string | NVD, AIID, or AIAAIC |
source_id |
string | Upstream identifier (CVE id, AIID id, AIAAIC id) |
title |
string | Short headline |
description |
string | Normalized description (PII masked) |
incident_date |
date | null | When the incident occurred |
ingested_at |
datetime | When this pipeline ingested the record |
vendor |
string | null | Alleged deployer / vendor |
system |
string | null | Product or system name |
atlas_technique |
string | MITRE ATLAS technique id or unmapped |
atlas_tactic |
string | null | ATLAS tactic name |
severity |
string | critical, high, medium, low, informational |
tags |
list[string] | Source-specific tags |
url |
string | null | Canonical reference URL |
raw |
object | Untouched upstream payload subset |
Sources
- NVD: AI/ML library CVEs (
pytorch,tensorflow,langchain,huggingface) - AIID: AI Incident Database
- AIAAIC: AIAAIC Repository
Usage
from datasets import load_dataset
ds = load_dataset("tk-tobi/ai-sentinel-feed")
print(ds["incidents"][0])
# Or load a single JSONL file directly:
ds = load_dataset("json", data_files="hf://datasets/tk-tobi/ai-sentinel-feed/incidents.jsonl")
Citation
If you use this dataset, cite the upstream sources (NVD, AIID, AIAAIC) and link to the ai-sentinel-feed repository.
License
Dataset compilation and normalization code is MIT-licensed. Upstream content remains subject to each source's terms.