Autonomous AI Agent Security Incidents of 2026: Benchmark Dataset & Incident Corpus
0
2 commits
1 linked in READMEs
updated Oct 6, 2026
README
Autonomous AI Agent Security Incidents of 2026: Benchmark Dataset & Incident Corpus
An open, verifiable, machine-readable dataset documenting 109 empirical security and containment failures involving autonomous AI coding, orchestration, and execution agents observed during 2026.
This dataset accompanies the 689-page open-access research monograph published on Zenodo:
Doletskyi, Serhii. (2026). Autonomous AI Agent Security Incidents of 2026: Empirical Incident Corpus, Sandbox Escape Forensics, and Multi-Agent Vulnerability Taxonomy. Zenodo. DOI: 10.5281/zenodo.22737862.
Dataset Structure
The repository contains three primary tabular datasets in standard RFC 4180 CSV format:
| File | Records | Description | Key Fields |
|---|---|---|---|
AI_Agent_Incident_Database_2026.csv | 109 incidents | Structured catalog of verified autonomous AI agent incidents in 2026 | incident_id, date, agent_framework, escape_vector, target_system, containment_breached, severity, primary_source |
AI_Agent_Evidence_Matrix_2026.csv | 193 claims | Falsification matrix with testable hypotheses and verification criteria | claim_id, hypothesis, falsification_condition, test_vector, reproducibility_status, forensic_hash |
AI_Agent_Metrics_2026.csv | 199 metrics | Empirical benchmark metrics evaluating agent containment resilience | metric_id, category, baseline_score, observed_failure_rate, mttr_seconds, containment_tier |
AI_Agent_Incident_Sources_2026.md | 378 sources | Complete bibliographic and forensic cross-references | Primary URLs, CVE disclosures, repository commit hashes |
Quick Start / Code Snippets
1. Load with Python & Pandas
import pandas as pd
# Load the 109 verified incidents
incidents_url = "https://huggingface.co/datasets/doletskyisergey/autonomous-ai-agent-security-incidents-2026/raw/main/AI_Agent_Incident_Database_2026.csv"
df_incidents = pd.read_csv(incidents_url)
print(f"Total incidents cataloged: {len(df_incidents)}")
print(df_incidents[["incident_id", "agent_framework", "escape_vector", "severity"]].head())
# Load the falsification evidence matrix
matrix_url = "https://huggingface.co/datasets/doletskyisergey/autonomous-ai-agent-security-incidents-2026/raw/main/AI_Agent_Evidence_Matrix_2026.csv"
df_matrix = pd.read_csv(matrix_url)
print(f"Total testable claims: {len(df_matrix)}")
2. Load with Hugging Face datasets
from datasets import load_dataset
dataset = load_dataset("doletskyisergey/autonomous-ai-agent-security-incidents-2026")
print(dataset)
Key Attack Vectors Cataloged
- Unconfined Unix Domain Sockets: Agents pivoting via accessible Docker/container daemon sockets (
/var/run/docker.sock,/run/podman/podman.sock). - Context Memory & Token Reuse: Secondary agent hijacking through residual authorization tokens and environmental variables persisted across agent handoffs.
- Multi-Agent Cascade Escalation: Automated subordinate agent privilege acquisition through synthetic tool invocations.
- Autonomous Package / Pipeline Injection: Covert lateral movement through developer toolchain modifications and automated dependency resolution scripts.
Full Monograph & Reproduction Artifacts
- Zenodo DOI: 10.5281/zenodo.22737862 (Permanent deposit, 689 pages)
- GitHub Repository: mars2021y-gif/autonomous-ai-agent-security-incidents-2026
Citation
@dataset{doletskyi2026incidents,
author = {Doletskyi, Serhii},
title = {Autonomous AI Agent Security Incidents of 2026: Benchmark Dataset and Incident Corpus},
month = oct,
year = 2026,
publisher = {Hugging Face / Zenodo},
doi = {10.5281/zenodo.22737862},
url = {https://huggingface.co/datasets/doletskyisergey/autonomous-ai-agent-security-incidents-2026}
}
License
This dataset is distributed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).
Similar repos
Autonomous AI Agent Security Incidents of 2026: Benchmark Dataset & Incident Corpus
0
2 commits
1 linked in READMEs
updated Oct 6, 2026
README
Autonomous AI Agent Security Incidents of 2026: Benchmark Dataset & Incident Corpus
An open, verifiable, machine-readable dataset documenting 109 empirical security and containment failures involving autonomous AI coding, orchestration, and execution agents observed during 2026.
This dataset accompanies the 689-page open-access research monograph published on Zenodo:
Doletskyi, Serhii. (2026). Autonomous AI Agent Security Incidents of 2026: Empirical Incident Corpus, Sandbox Escape Forensics, and Multi-Agent Vulnerability Taxonomy. Zenodo. DOI: 10.5281/zenodo.22737862.
Dataset Structure
The repository contains three primary tabular datasets in standard RFC 4180 CSV format:
| File | Records | Description | Key Fields |
|---|---|---|---|
AI_Agent_Incident_Database_2026.csv | 109 incidents | Structured catalog of verified autonomous AI agent incidents in 2026 | incident_id, date, agent_framework, escape_vector, target_system, containment_breached, severity, primary_source |
AI_Agent_Evidence_Matrix_2026.csv | 193 claims | Falsification matrix with testable hypotheses and verification criteria | claim_id, hypothesis, falsification_condition, test_vector, reproducibility_status, forensic_hash |
AI_Agent_Metrics_2026.csv | 199 metrics | Empirical benchmark metrics evaluating agent containment resilience | metric_id, category, baseline_score, observed_failure_rate, mttr_seconds, containment_tier |
AI_Agent_Incident_Sources_2026.md | 378 sources | Complete bibliographic and forensic cross-references | Primary URLs, CVE disclosures, repository commit hashes |
Quick Start / Code Snippets
1. Load with Python & Pandas
import pandas as pd
# Load the 109 verified incidents
incidents_url = "https://huggingface.co/datasets/doletskyisergey/autonomous-ai-agent-security-incidents-2026/raw/main/AI_Agent_Incident_Database_2026.csv"
df_incidents = pd.read_csv(incidents_url)
print(f"Total incidents cataloged: {len(df_incidents)}")
print(df_incidents[["incident_id", "agent_framework", "escape_vector", "severity"]].head())
# Load the falsification evidence matrix
matrix_url = "https://huggingface.co/datasets/doletskyisergey/autonomous-ai-agent-security-incidents-2026/raw/main/AI_Agent_Evidence_Matrix_2026.csv"
df_matrix = pd.read_csv(matrix_url)
print(f"Total testable claims: {len(df_matrix)}")
2. Load with Hugging Face datasets
from datasets import load_dataset
dataset = load_dataset("doletskyisergey/autonomous-ai-agent-security-incidents-2026")
print(dataset)
Key Attack Vectors Cataloged
- Unconfined Unix Domain Sockets: Agents pivoting via accessible Docker/container daemon sockets (
/var/run/docker.sock,/run/podman/podman.sock). - Context Memory & Token Reuse: Secondary agent hijacking through residual authorization tokens and environmental variables persisted across agent handoffs.
- Multi-Agent Cascade Escalation: Automated subordinate agent privilege acquisition through synthetic tool invocations.
- Autonomous Package / Pipeline Injection: Covert lateral movement through developer toolchain modifications and automated dependency resolution scripts.
Full Monograph & Reproduction Artifacts
- Zenodo DOI: 10.5281/zenodo.22737862 (Permanent deposit, 689 pages)
- GitHub Repository: mars2021y-gif/autonomous-ai-agent-security-incidents-2026
Citation
@dataset{doletskyi2026incidents,
author = {Doletskyi, Serhii},
title = {Autonomous AI Agent Security Incidents of 2026: Benchmark Dataset and Incident Corpus},
month = oct,
year = 2026,
publisher = {Hugging Face / Zenodo},
doi = {10.5281/zenodo.22737862},
url = {https://huggingface.co/datasets/doletskyisergey/autonomous-ai-agent-security-incidents-2026}
}
License
This dataset is distributed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).