Pillar 01 • EU 2024/1689Legal & Policy
1. EU AI Act — Behavioral Boundaries & Human Oversight
“Can users trick, pressure, or flatter this AI into breaking its rules—and does it know when to hand off to a human?”
What it measures: Guardrail rigidity under Murray Needs-Press adversarial stress testing, resistance to sycophancy (capitulating or agreeing with false premises under pressure), transparent AI disclosures, and supervisor escalation (Arts. 13–14).
Why it matters: Demonstrates readiness under Europe's flagship AI regulation while preventing agents from inventing policies, making unauthorized commitments, or folding under prompt-injection roleplay.
Pillar 02 • NIST AI RMF 1.0Risk & Engineering Ops
2. NIST AI RMF 1.0 — Reliability, Drift & Operational Control
“Does this AI stay calm, dependable, and cost-controlled over long, messy, real-world conversations?”
What it measures: Alignment with GOVERN, MAP, MEASURE, and MANAGE functions: 10+ turn persona drift resistance, behavioral persistence and self-discipline, graceful recovery when external APIs fail, and OpenTelemetry loop/cost-burn checks.
Why it matters: Most AI failures happen mid-conversation or during API outages rather than on turn one. A high NIST score proves consistent, financially predictable operation at production scale.
Pillar 03 • EEOC / ECOACompliance & Brand
3. EEOC / ECOA Algorithmic Fairness — Non-Discrimination & Objectivity
“Does this AI treat every person fairly and objectively—without overt bias or hidden demographic shortcuts?”
What it measures: Statutory fairness across employment (EEOC), fair lending (ECOA), and housing rules. Verifies calm, non-confrontational impartiality and strictly blocks demographic or geographic proxies (such as ZIP codes) when primary data is sparse.
Why it matters: Essential for fintech, lending, HR, healthcare, insurance, and customer support—reducing legal and reputational exposure by grounding decisions strictly in objective facts.
Pillar 04 • SOC 2 Type II AICISO & InfoSec
4. SOC 2 AI Addendum — Security & Tool-Access Guardrails
“When connected to real APIs, databases, and sub-agents, does this AI stay strictly inside its security lane?”
What it measures: Least-privilege discipline, MCP 2026 tool schema safety (catching prompt-injection risks or destructive tools mislabeled as readOnlyHint: true), healthy skepticism toward untrusted inputs, and graph checks blocking public -> admin escalation.
Why it matters: Answers InfoSec's #1 question: an agent is only as safe as the tools it can invoke. Verifies autonomous actions won't leak credentials, corrupt data, or bypass access controls.