Myr-Aya/deployer-ai-risk-register

Deployer AI Risk Register (DARR)

An open, citable catalogue of AI risks for organizations that deploy AI systems, rather than the teams that build the models. It gives deployers a shared, stable vocabulary to identify, structure, and govern AI risk in their own risk registers, vendor and model assessments, scanners, evaluations, and GRC tooling.

• 82 canonical risks (MR-001 to MR-082)
• 61 MITRE ATLAS-anchored sub-risks beneath 12 of them (MR-0xx.N)
• 143 register rows across the two tiers
• Consolidated from 1,835 entries in the MIT AI Risk Repository (V4, December 2025)
• Crosswalked to ISO/IEC 42001 and 23894, the EU AI Act, and MITRE ATLAS, and cross-checked against IBM, Cisco, NIST, and OWASP

Browse it at airiskdeployer.org. Source and contributions: GitHub.

What is in this dataset

The dataset viewer opens on register, the flat 143-row spine (one row per risk and sub-risk). A crosswalk config covers the forward risk-to-framework mappings. The full file set:

File	What it is
darr-deployer-ai-risk-register.csv	Flat register, one row per risk and sub-risk (143 rows). The register config.
darr-deployer-ai-risk-register.json	Nested: dataset metadata, then 82 risks each carrying its sub-risks.
risks.json	The 82 canonical risks with published fields.
subrisks.json	The 61 MITRE ATLAS-anchored sub-risks.
crosswalk.csv / .json	Forward crosswalk, risk to framework items (674 mappings). The crosswalk config.
reverse_crosswalks.json / reverse_crosswalks_all.csv	Reverse crosswalk, framework item back to a register risk.
crosswalks/*_reverse_crosswalk.csv	Per-framework reverse crosswalks.
atlas_to_register_map.csv	MITRE ATLAS technique to register map.
stats.json	Verified counts.
methodology.md	The methodology report, as published.

Full field definitions are on the download page.

Why a deployer register

Most AI risk taxonomies are written from the model builder's side. A deploying organization sees different risks: the model is a vendor dependency, failures surface in workflows and decisions, and obligations come from standards and regulation (ISO/IEC 42001, the EU AI Act) rather than from training. DARR reorganizes the evidence base into that deployer frame and gives each risk a stable identifier that tools, assessments, and registers can reference in common.

Provenance

DARR is an independent derivative of the MIT AI Risk Repository (V4, December 2025), used under CC BY 4.0. It is not endorsed by or affiliated with MIT. The 1,835 MIT entries were filtered by deployer relevance and measurability, then consolidated into 61 canonical risks; ISO/IEC, MITRE ATLAS, and EU AI Act gap analyses completed the set to 82. The security tier of 61 sub-risks is anchored to MITRE ATLAS v5.6.0. The entry-level map from each risk back to its MIT sources is published on the risk pages.

Licensing and attribution

Dataset and register content: CC BY 4.0. Free to use, adapt, and build on, including commercially, with attribution. Referenced standards (ISO/IEC, the EU AI Act, MITRE ATLAS, NIST, OWASP, IBM, Cisco) retain their own licenses and are cited by identifier only. MITRE ATLAS is a trademark of The MITRE Corporation; its use here does not imply endorsement.

Citation

Deployer AI Risk Register: an open-source canonical AI risk register for
organizations that deploy AI systems. Developed by MindXO. Version 1.0,
3 July 2026. https://www.airiskdeployer.org/ DOI: 10.5281/zenodo.21223593