🤖NIST AI-RMF MEASURE-2.6Rule: AIRMF-ME-003high

Trustworthiness characteristics evaluated and documented

Description

Validity, reliability, safety, security, resilience, accountability, transparency, explainability, privacy, and fairness are all evaluated and reported.

⚠️ Risk Impact

Optimising only one trustworthiness dimension (e.g. accuracy) without measuring the others produces models that fail in the unmeasured dimensions. The Samsung ChatGPT leak (April 2023) is a privacy-dimension failure that wouldn't have shown in accuracy metrics.

🔍 How EchelonGraph Detects This

AIRMF-ME-003Automated scanner rule

EchelonGraph's Tier 1 Cloud Scanner automatically checks for this condition across all connected cloud accounts. Violations are flagged as high-severity findings with remediation guidance.

🔧 Remediation

Maintain a trustworthiness dashboard with each AI-RMF dimension scored or RAG-rated per model. Update on release; brief leadership quarterly.

💀 Real-World Attack Scenario

Samsung Semiconductor engineers pasted internal source code into ChatGPT to debug a defect (April 2023). OpenAI's training pipeline absorbed the data; the snippets were potentially recoverable through prompt completion by competitors. Samsung banned ChatGPT internally within 7 days; estimated IP exposure included proprietary semiconductor designs worth an undisclosed multi-million-dollar amount.

💰 Cost of Non-Compliance

Samsung ChatGPT leak (Apr 2023): undisclosed but material IP exposure. Enterprise AI data-loss incidents in 2024: avg $3.8M per occurrence (IBM). EU AI Act Article 15 cybersecurity: €15M / 3% revenue.

📋 Audit Questions

  • 1.Show me the trustworthiness scorecard for your top 3 deployed models.
  • 2.Which dimensions scored 'red' in the last quarter? What action followed?
  • 3.How is the scorecard surfaced to non-technical stakeholders?
  • 4.What is your data-leak prevention process for staff using third-party LLMs?

🎯 MITRE ATT&CK Mapping

T1530 — Data from Cloud StorageT1213.003 — Code Repositories

⚡ Common Pitfalls

  • Scoring only the easy dimensions (accuracy, fairness) and skipping the harder ones (explainability, accountability)
  • Letting the trustworthiness scorecard fall stale (>90 days old) — auditors view stale as worse than absent
  • No alerting on dimension-level regression — the scorecard is purely a snapshot

📈 Business Value

Multi-dimensional trustworthiness assessment + alerting catches failures across the long tail of trustworthiness — privacy leaks, robustness regressions, fairness drift — that single-metric monitoring misses.

⏱️ Effort Estimate

Manual

2-4 weeks initial scorecard build per model

With EchelonGraph

EchelonGraph ships scorecards per workload with auto-updated dimension scores from telemetry + eval

🔗 Cross-Framework References

EU_AI_ACT-ART15-CYBERSECISO42001-9.1

Automate NIST AI-RMF MEASURE-2.6 compliance

EchelonGraph continuously monitors this control across all your cloud accounts.

Start Free →