BibTeX Citations
Use the following entries to cite Failure-First research in academic work. Click any block to copy.
Framework
@misc{failurefirst2025framework,
title = {Failure-First Embodied AI: A Framework for
Characterizing Adversarial Failure in
Embodied AI Systems},
author = {Wedd, Adrian},
year = {2025},
url = {https://failurefirst.org},
note = {Version 0.13, dataset v0.2}
}Dataset
@misc{failurefirst2025dataset,
title = {Failure-First Embodied AI Adversarial
Scenario Dataset},
author = {Wedd, Adrian},
year = {2025},
url = {https://github.com/adrianwedd/failure-first},
note = {142,307+ scenarios, 661 failure classes,
19 domains, JSONL format}
}Methodology
@misc{failurefirst2026methodology,
title = {Adversarial Evaluation Methodology for
Embodied AI Safety},
author = {Wedd, Adrian},
year = {2026},
url = {https://failurefirst.org/research/methodology/},
note = {Multi-phase evaluation: scenario construction,
multi-model evaluation, failure classification}
}Moltbook Research
@misc{failurefirst2026moltbook,
title = {Multi-Agent Attack Surface Analysis:
Empirical Study of AI Agent Interactions
on Moltbook},
author = {Wedd, Adrian},
year = {2026},
url = {https://failurefirst.org/research/moltbook/},
note = {1,497 posts classified against 34+ attack
patterns using regex and LLM semantic analysis}
}Data Access
Public Data
The following are freely available:
- JSON Schemas for all dataset formats (single-agent, multi-agent, episode)
- Attack taxonomy with 346+ pattern categories and descriptions
- Failure mode taxonomy with recursive failure classifications
- Recovery mechanism taxonomy
- Benchmark pack configurations (YAML)
- Evaluation tools (validators, linters, benchmark runners)
- Aggregate results and metrics (this site)
Research Data (By Request)
The following require a research data access request. This data is maintained in a private repository to prevent misuse of operational attack content:
- Full adversarial scenario datasets (JSONL with specific prompts)
- Model evaluation traces (per-scenario input/output)
- Moltbook corpus with classified posts
- Compression tournament results with specific prompts
- Multi-agent scenario scripts with full actor dialogues
To request access, contact research@failurefirst.org with your institutional affiliation and intended use.
Public Metadata
Machine-readable metadata for the dataset and research program:
Dataset Summary (v0.2)
Responsible Disclosure
If you discover a vulnerability in a deployed AI system using insights from this research, please follow responsible disclosure practices. See our responsible disclosure page for guidance.
License
The Failure-First framework, tools, and public documentation are released under the MIT License. Research data access is granted on a case-by-case basis for legitimate AI safety research purposes.