NIST AI Risk Management Framework for LLM deployments
The NIST AI Risk Management Framework (AI RMF 1.0) is the U.S. federal government's voluntary framework for managing AI risk. It is increasingly referenced by procurement requirements, cyber insurance underwriters, and as the technical vocabulary in regulatory guidance. For fintech firms, it complements SR 11-7 and EU AI Act requirements.
The four core functions
NIST AI RMF is organized into four functions. LLMSec primarily addresses the MEASURE function, which covers quantifying and testing AI risk — but the scan outputs also support GOVERN and MANAGE.
Policies, culture, accountability. Sets the organizational context for AI risk management.
LLMSec's audit trail and signed reports provide evidence for governance processes.
Categorize AI systems and identify risks in context.
LLMSec probe categories map to NIST's trustworthy characteristics taxonomy.
Quantify and evaluate AI risk. This is where systematic testing occurs.
LLMSec's primary function — 30+ probes, severity classification, documented methodology.
Prioritize and address identified risks. Track risk posture over time.
LLMSec's remediation guidance and re-scan capability support ongoing risk management.
MEASURE function — subcategory mapping
NIST AI RMF subcategories within MEASURE that LLMSec directly addresses:
MEASURE 2.5: AI risks and benefits of AI systems
LLMSec's severity classification (CRITICAL/HIGH/MEDIUM/LOW) provides a quantified risk posture per probe category, giving the basis for MEASURE 2.5 risk documentation.
MEASURE 2.6: Testing on a representative sample
LLMSec runs 30+ diverse probes across jailbreak, injection, toxicity, data extraction, and reliability categories — constituting a representative adversarial sample for each deployed model.
MEASURE 2.7: AI system to be monitored for trustworthy characteristics
Probe categories directly map to NIST AI RMF trustworthy characteristics: Reliable (reliability probes), Safe (toxicity, jailbreak), Secure (injection, data extraction), Explainable (refusal consistency).
MEASURE 2.10: Privacy risks of AI system
Data extraction and PII leakage probes test whether the model exposes training data, system prompts, or personally identifiable information.
MEASURE 2.11: Fairness and bias
Toxicity probes include bias detection across protected characteristics. Discriminatory output testing documents model behavior across demographic groups.
MEASURE 4.1: AI risks and impacts are measured, assessed, and documented
Every LLMSec scan produces a documented, signed report with probe-level findings, severity assessments, and framework mappings — the core deliverable for MEASURE 4.1.
NIST trustworthy AI characteristics
NIST AI RMF defines seven characteristics of trustworthy AI. LLMSec probe categories map directly to these characteristics, making it easy to show which characteristics your model meets and which have identified gaps:
| NIST Characteristic | LLMSec Probe Categories |
|---|---|
| Reliable | Reliability, refusal consistency, multi-turn coherence |
| Safe | Jailbreak, role-play bypass, constraint override |
| Secure & Resilient | Prompt injection, system prompt leakage, encoding bypass |
| Explainable & Interpretable | Refusal reasoning, output consistency |
| Privacy-Enhanced | PII leakage, training data extraction, system prompt leakage |
| Fair with Harmful Bias Managed | Toxicity, hate speech, discriminatory output |
| Accountable & Transparent | Audit trail, signed scan report, scan fingerprint |
Why NIST AI RMF is showing up in procurement
The 2023 Executive Order on AI required federal agencies to use NIST AI RMF as their AI risk management framework. This has cascaded into procurement requirements: federal contractors and regulated entities serving government-adjacent clients are increasingly required to demonstrate AI RMF alignment.
Cyber insurance underwriters are beginning to ask about AI risk management as part of policy applications. NIST AI RMF is the vocabulary they use. An LLMSec report mapped to NIST provides concrete evidence for these inquiries.
Finally, NIST AI RMF and FINRA SR 11-7 are complementary: SR 11-7 provides the financial-services-specific model risk governance requirements; NIST AI RMF provides the AI-specific technical testing vocabulary. An LLMSec report maps findings to both simultaneously.
How is accuracy measured?The MEASURE function of NIST AI RMF requires documented evaluation methodology. Every LLMSec judge's precision, recall, F1, and false-negative rate is published — including weaknesses — on the calibration & methodology page.