NIST AI 800-1 2pd Managing Misuse Risk for Dual-Use Foundation Models
- Publisher: NIST, U.S. AISI
- Status:
draft - Version:
2pd - Release Date:
2025-01-15 - Date Added:
2025-09-30 - Source URL: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.800-1.ipd2.pdf
Summary
This draft guideline provides voluntary, non-regulatory recommendations for how organizations (especially foundation model developers) can anticipate, measure, and mitigate the risk that their models might be deliberately or indirectly misused. It is targeted at the “initial developer” in the AI supply chain but also offers guidance for other stakeholders. The guidance spans the AI lifecycle, from threat modeling and evaluation, through deployment, monitoring, response, and disclosure, particularly in domains with high public safety or national security implications.
Key Takeaways
- Organized recommendations around seven objectives: identify potential misuse, plan, protect, measure, mitigate, monitor/respond, disclose.
- Red-teaming and misuse evaluations are expected practices for every foundation model
- Addresses some domain-specific misuse risks in the appendix; chemical/biological and cybersecurity
- Transparency & disclosure are essential parts of responsible release
Additional Sources
Tags
llm, frontier-model, misuse, transparency
License
Public-domain