Dataset · Text classification
ai-agent-security-incidents
by Legion 0 gemmozero/ai-agent-security-incidents
A structured, machine-readable database of 1330 confirmed AI agent security incidents, collected and classified automatically.
Dataset Card
By Legion 0, published under cc-by-4.0, revision b41c239cee3b.
AI Agent Security Incident Database v0.1
A structured, machine-readable database of 1330 confirmed AI agent security incidents, collected and classified automatically.
What is this?
Every time an AI agent causes unintended harm — escaping a sandbox, exploiting an API, taking unauthorized actions, exfiltrating data — this database captures it.
This is not a list of theoretical risks. Every entry describes something that actually happened, with a verifiable source URL.
Schema
| Field | Type | Description |
|---|---|---|
| id | integer | Unique identifier |
| source | string | Collection source (hackernews, wired, cve_nvd, etc.) |
| source_url | string | Verifiable URL |
| title | string | Incident title |
| date_incident | string | YYYY-MM-DD or "unknown" |
| date_collected | string | ISO 8601 |
| classifier_model | string | Model used for classification |
| faille_type | string | See categories below |
| severity | string | low, medium, high, critical |
| framework | string | AI framework involved or "unknown" |
| entreprise | string | Company involved or "unknown" |
| api_involved | integer | 1 if API was involved |
Incident Categories
| Category | Description |
|---|---|
| api_exploit | Agent discovered and exploited API vulnerabilities |
| agent_escape | Agent broke out of its designated environment |
| sandbox_escape | Agent escaped sandboxed/containerized environment |
| unauthorized_action | Agent acted beyond its mandate |
| data_exfiltration | Agent leaked or stole sensitive data |
| prompt_injection | Agent was manipulated via prompt injection |
| security_by_obscurity | Agent bypassed security that relied on obscurity |
| supply_chain | Attack via AI supply chain (packages, models) |
| model_manipulation | Model weights or training data were compromised |
| configuration_exploit | Agent exploited misconfiguration |
Sources
Data is collected daily from 30+ public sources including: - CVE/NVD (National Vulnerability Database) - GitHub Security Advisories - Wired, Ars Technica, The Hacker News, SecurityWeek - Schneier on Security, Krebs on Security - DarkReading, BleepingComputer - HackerNews - arXiv security papers
Methodology
- Automated collection from public sources via APIs and RSS feeds
- AI classification using Llama 3.1 8B (local, on-device)
- Structured extraction: incident type, severity, cause chain
- Normalization: single-value types and severities, no composites
- Quality filter: entries without verifiable source or description are rejected
Limitations
- Classification is automated (8B parameter model) — expect some errors
- Some entries may be misclassified demonstrations or research
- Severity is model-assigned, not manually verified
- Coverage starts August 2026
- This is v0.1 — the schema may evolve
Files
ai_agent_security_incidents_v0.1.json— Full dataset with metadata and statisticsai_agent_security_incidents_v0.1.csv— Flat CSV for easy analysis
Updates
This database is updated daily via automated collection. New versions will be published periodically.
Live API
curl "https://api.legion-api.com/incidents?limit=10" # free: 10 results, metadata only
curl "https://api.legion-api.com/incidents?type=prompt_injection"
curl "https://api.legion-api.com/incidents/stats"
curl "https://api.legion-api.com/incidents?since=<id>" # cursor pagination
Full access — 500 results/call, descriptions, root-cause chains, daily JSON export: API key €29/month · Archive €299 · MCP server · Docs
Related datasets (same pipeline, updated daily)
ai-models-2026 · ai-policydrift-2026 · ai-compute-prices-2026 · ToS history 2023→2026
Premium
Full descriptions, root-cause chains and the real-time feed: API · Gumroad
License
CC BY 4.0 — Free to use with attribution.
Contact
Issues and feedback welcome via GitHub issues or HuggingFace discussions.
Structure
default 1,430 rows
| Split | Rows | Size |
|---|---|---|
| train | 1,430 | 412.1 KB |
Details
- Repository
- gemmozero/ai-agent-security-incidents
- Publisher
- Legion 0
- Task category
- Text classification
- Tags
- ai-safety, security, ai-agents
- Size category
- Not stated by the source
- Languages
- ai-safety, ai-agents
- Revision
- b41c239cee3b8e6353040e9a456db87ffb64cc02
- Last updated
- 2026-09-20
Files
4 files, 369.9 KB in total.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| ai_agent_security_incidents_v0.1.csv | Data | 341.3 KB | — |
| sample_100_free.csv | Data | 21.4 KB | — |
| README.md | Documentation | 4.7 KB | — |
| .gitattributes | Repository | 2.5 KB | — |
License and Download
- License
- cc-by-4.0
- Access
- No access gate
Released by Legion 0 through its official repository on Hugging Face. Read the license.