Glossary
NIST AI RMF
A voluntary framework from the US National Institute of Standards and Technology for managing AI risk across the life cycle: four functions, seven trustworthiness characteristics, and a 2024 profile for generative AI.
General definition
The NIST AI RMF (Artificial Intelligence Risk Management Framework, document NIST AI 100-1) was released on 26 January 2023 after a public drafting process that the US National AI Initiative Act of 2020 asked NIST to run. It is voluntary, sector-neutral and free to use, and it has become the common vocabulary for AI risk in the United States much as the NIST Cybersecurity Framework did for security. It defines trustworthy AI by seven characteristics: valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; and fair, with harmful bias managed.
The core of the framework is four functions, each broken into categories and subcategories of outcomes:
- Govern: the cross-cutting function: policies, roles, accountability, culture and processes for the whole life cycle, including third-party and supply-chain risk
- Map: establish the context: what the system is for, who it affects, what could go wrong and whether the benefits justify the risk at all
- Measure: test and evaluate the identified risks with quantitative and qualitative methods, track them over time and check that the metrics themselves are sound
- Manage: prioritise and act on the risks: accept, mitigate, transfer or avoid them, respond to incidents and plan for decommissioning
NIST publishes a companion Playbook with suggested actions for every subcategory and encourages organisations to write profiles: a tailored selection of outcomes for a sector or use case. The most important is the Generative AI Profile (NIST AI 600-1, July 2024), written under the October 2023 Executive Order on AI. It names twelve risks that are new or amplified in generative AI, including confabulation (hallucination), data privacy, information integrity, information security, intellectual property, harmful bias, human-AI configuration and value-chain integration, and lists suggested actions for each, mapped to the four functions.
Because it is voluntary, the AI RMF is most often used as an internal yardstick and as a bridge to other regimes: it maps closely onto ISO/IEC 42001, US agencies reference it in procurement, and companies use it to structure the technical file an EU AI Act high-risk system needs. Its main weakness is also its strength: it says what outcomes to reach, not how, so two organisations can both claim alignment while doing very different amounts of work.
In the Ethora ecosystem
Mapped to the four functions, Ethora covers the parts a platform can. For Govern and Map, every AI agent has a documented identity: a name, system prompt, model, knowledge base and tool list in the admin panel, so the inventory and the intended-purpose description exist from day one. For Measure, agents can be evaluated before release against a fixed model version, with LLM evaluation sets and red teaming run against the same configuration users will see; because the model endpoint is chosen by you, the version under test is the version in production.
For Manage, the controls live in the messaging layer: the Trust and Safety system can block, censor or quarantine what an agent sends, multi-agent rooms have a response gate and an autonomous-turn cap, and every exchange is in the archive and the compliance audit trail. The Generative AI Profile’s data privacy and information security risks are addressed structurally on a dedicated or self-hosted deployment, where prompts, transcripts and knowledge stay inside your own perimeter and can be served by a self-hosted model with no third-party API in the loop. Treat this as alignment with the framework rather than certification: the AI RMF has no certification scheme, and AI governance remains your programme.