NIST SP 1353: using AI with CSF 2.0 without turning a draft into an auditor

The signal and the facts that change the decision

NIST published the SP 1353 initial public draft on 19 August, with comments open until 15 October 2026. The guide provides structured prompts, simulated documents, and three use cases: governance review, building a Current State Profile, and designing a Target State Profile aligned with CSF 2.0 outcomes.

NIST itself labels the publication an Initial Public Draft and says its focus is not general AI best practice or complete cybersecurity guidance. The three use cases rely on notional material from a fictitious organisation; outputs generated from that corpus do not validate performance in a real enterprise.

CSF 2.0 organises outcomes and profiles, while AI RMF addresses risks from using AI; the two frameworks have complementary roles and should not be collapsed into a single score. A model can produce a coherent profile while inventing controls, confusing policy with execution, omitting contradictory interviews, or converting missing evidence into compliance; documentary appearance is not assurance.

Decision, limits, and stop condition

Use the prompts as analysis accelerators, never as evidence sources: load only authorised artefacts, require references per claim, mark assumptions and gaps, and make a human owner accept or reject every mapping.

A model can produce a coherent profile while inventing controls, confusing policy with execution, omitting contradictory interviews, or converting missing evidence into compliance; documentary appearance is not assurance.

Execution and proof of completion

Execution should start by selecting a narrow domain and creating a gold set of policies, evidence, interviews, gaps, and approved mappings. Then, running the prompts with mandatory citations and measuring mapping accuracy, hallucination, coverage, time, and human disagreement. Closure requires publishing only the reviewed profile while retaining model version, prompt, inputs, outputs, corrections, and final decision.

The pilot passes when no outcome is marked achieved without traceable evidence, every unknown gap remains open, and a second person reproduces the decision from the retained package.

• selecting a narrow domain and creating a gold set of policies, evidence, interviews, gaps, and approved mappings.
• running the prompts with mandatory citations and measuring mapping accuracy, hallucination, coverage, time, and human disagreement.
• publishing only the reviewed profile while retaining model version, prompt, inputs, outputs, corrections, and final decision.
• The pilot passes when no outcome is marked achieved without traceable evidence, every unknown gap remains open, and a second person reproduces the decision from the retained package.