NIST's Generative AI Profile, AI 600-1, lists 12 risks that are new or made worse by generative AI. For a RAG assistant, test data privacy, information security, confabulation, and value chain and component integration first, because retrieval touches all four.
Where the profile sits
NIST released the Generative AI Profile on 26 July 2024 as a companion to the AI Risk Management Framework, which NIST published in its 1.0 form on 26 January 2023. The profile helps organisations identify risks specific to generative AI and proposes actions that fit their goals. NIST's AI RMF page says the framework is being revised as part of the White House AI Action Plan, so check for updates.
The 12 risks
- CBRN Information or Capabilities
- Confabulation
- Dangerous, Violent, or Hateful Content
- Data Privacy
- Environmental Impacts
- Harmful Bias or Homogenization
- Human-AI Configuration
- Information Integrity
- Information Security
- Intellectual Property
- Obscene, Degrading, and/or Abusive Content
- Value Chain and Component Integration
Which matter first for retrieval
Data Privacy. NIST describes impacts from leakage and unauthorised use, disclosure or de-anonymisation of personal data. A RAG index full of internal documents is the obvious path. See permission-aware retrieval.
Information Security. NIST describes lowered barriers for offensive cyber capabilities and an increased attack surface that may compromise a system's availability or the confidentiality or integrity of training data, code or model weights. Poisoned content and prompt injection fit the second part.
Confabulation. Confidently stated but wrong content. Retrieval can reduce it, but a poisoned or outdated source makes the assistant confidently wrong with a citation.
Value Chain and Component Integration. NIST describes non-transparent or untraceable integration of upstream third-party components. A RAG stack has many: embedding models, vector databases, loaders, rerankers, document sources. Knowing where each came from is the control.
Human-AI Configuration. NIST lists automation bias and over-reliance among the outcomes. If users treat retrieved answers as verified, a poisoned source gets more trust than it should.
Lower priority, still check
Intellectual Property matters if you index licensed or third-party content. Information Integrity matters if the assistant produces content people will publish. Harmful content risks matter mostly for public-facing systems.
A one-page use
Make a table with the 12 risks in rows and your assistant's components in columns. In each cell write "test", "accept" or "not applicable", with a reason. Share it with product and legal. It becomes your test plan and a record of decisions.
NIST's profile also suggests actions such as AI red-teaming and structured public feedback as ways to find these risks before and after release. Pair this with your red team scoping.
Quick check
Is the Generative AI Profile a regulation?
No. It is voluntary NIST guidance that companions the AI Risk Management Framework.
Do I need to address all 12 risks equally?
No. Rank them by your use case and record why you accepted or deferred the rest.
Sources
Last reviewed: 19 October 2026.
