NIST AI 600-1, the Generative AI Profile released on 26 July 2024, lists twelve risks that generative AI creates or makes worse: CBRN information, confabulation, harmful content, data privacy, environmental impacts, bias, human-AI configuration, information integrity, information security, intellectual property, abusive content, and value chain integration.
The twelve, in plain words
- CBRN information or capabilities. Easier access to dangerous chemical, biological, radiological or nuclear information. Mostly a concern for model builders.
- Confabulation. Confidently stated but false content, often called hallucination. Matters whenever someone acts on output.
- Dangerous, violent or hateful content. Easier production of threatening or inciting content.
- Data privacy. Leakage, unauthorized use or de-anonymization of personal or sensitive data.
- Environmental impacts. High compute use in training or operating models.
- Harmful bias or homogenization. Amplified biases, performance gaps between groups, and skewed uniform output.
- Human-AI configuration. Over-reliance, automation bias, anthropomorphizing, or emotional entanglement.
- Information integrity. Lower barriers to producing content that blurs fact and fiction at scale.
- Information security. Lower barriers for offensive cyber operations, and a larger attack surface for training data, code and model weights.
- Intellectual property. Easier replication of copyrighted or licensed content and exposure of trade secrets.
- Obscene, degrading or abusive content. Includes synthetic abusive imagery.
- Value chain and component integration. Untraceable upstream components, improper supplier vetting, and reduced transparency for downstream users.
Which ones matter for a typical software company
Most product teams will spend their time on confabulation, data privacy, information security, human-AI configuration, intellectual property, and value chain integration. The others still belong in a risk review, but often as "not applicable, because" rather than as active work.
How to use the list
Add a column to your inventory for "applicable GAI risks." During each system's assessment, walk all twelve and mark yes, no or partial. For each yes, name a control and a test.
The profile organizes suggested actions by the same GOVERN, MAP, MEASURE and MANAGE subcategories as the main framework, and tags each action with the risks it addresses. That makes it easy to filter: pick a risk, then read the actions tagged with it.
What it is not
It is a voluntary profile, not a legal requirement and not a certification. NIST's overview describes it as helping organizations identify risks unique to generative AI and proposing actions aligned to their goals.
Common misreadings
Some teams treat the twelve as a ranking. NIST does not rank them, and the weight each carries depends on your use case. Others assume that a vendor has handled them. A hosted model provider may address some, but you still own how you deploy, what data you send and how people use the output. Finally, the list is about generative AI specifically, so classic machine learning systems still need the main framework's attention to bias, validity and monitoring.
Quick check
Are these risks only for companies that train models?
No. Several, such as confabulation, data privacy and value chain integration, apply to anyone who deploys or uses generative AI.
Is "hallucination" the same as confabulation?
NIST uses confabulation and notes that hallucination is the colloquial term. It also notes that anthropomorphic terms can themselves feed over-reliance.
Sources
Last reviewed: 19 October 2026.
