NIST AI RMF — satisfying MEASURE 2.7 with Provael

NIST AI RMF organises trustworthiness into GOVERN / MAP / MEASURE / MANAGE. MEASURE 2.7 is "AI system security and resilience" — it expects evaluated adversarial/red-team testing, not an assertion. A Provael run is a clean fit: a controlled, reproducible measurement that feeds MANAGE.

Walkthrough

provael calibrate --policy stub --suite stub --seeds 20 --out calib
provael attack --policy stub --suite stub --calib calib \
    --attacks none,instruction,visual,injection,action --episodes 10 --out runs/measure
provael report --in runs/measure --format oscal --out report.oscal.json   # GRC-ingestible

How it maps

AI RMF function	What Provael provides
GOVERN / MAP	The Embodied AI Security Top-10 (EAI01–10) as the mapped risk context (report.json#/eai)
MEASURE 2.7 (security & resilience)	Calibrated redirection rate + 95% Wilson CI + benign-FPR control — a measured, controlled metric, not a vibe
MANAGE	Re-run per checkpoint; track ASR over time (leaderboard); regression-gate (examples/ci/regression-gate.md)

Tag findings with MITRE ATLAS technique ids and the NIST AI 100-2e2025 adversarial-ML classes (evasion / poisoning / privacy) for cross-framework traceability — see COMPLIANCE.md and the OSCAL export (observations carry the EAI id; map those to ATLAS in your GRC tool).

Evidence, not certification. n = 10 per attack is a screen — read the CIs, run more seeds for a stronger claim, and always read each rate against the benign control.