← Accueil actu | Salle du marché
2026-09-14 · Source:ITWorld Korea

Generative AI programs stay stuck in pilot without engineered governance, argues Adnan Masood — policy must become executable rules, evidence and drift measurement.
Governance determines whether generative AI programs remain stuck in pilot or evolve into a lasting capability, Adnan Masood argues in an ITWorld Korea analysis. He treats governance as engineering work: embed policy into build-time and runtime controls that control access, restrict behavior, collect evidence and measure drift — a concept he calls governance by design. The goal is simple: systems must operate within policy boundaries daily even as conditions change. Policy becomes executable rules, evidence is generated as a byproduct of normal operations, and governance running inside the deployment pipeline speeds up delivery.
Enterprises already know the pattern: payment systems carry built-in fraud and chargeback prevention, while customer data platforms embed consent and data-retention controls. Generative AI needs the same treatment because it increasingly crosses data boundaries, generates content and acts directly through tools.
The starting point is a threat model grounded in actual usage — a concise account of what can go wrong, who is harmed and where controls belong — kept concrete and focused on production failures. Masood enumerates five: data exposure through prompts, logs or model outputs, letting sensitive data cross trust boundaries quickly; prompt injection and indirect instructions, where user messages or retrieved documents can induce unsafe behavior; source integrity risks, where seemingly benign but stale, inaccurate or tampered documents surface via retrieval; tool misuse, since agents and assistants can write to systems of record through APIs and workflow tools; and model and prompt drift, as systems change through model updates, prompt edits and data refresh cycles.
Controls should follow that threat model. Masood rejects confining them to process stages alone; runtime and CI controls lessen reliance on memory and manual review.
The approach rests on standardized governance primitives — the smallest units implementable and reusable across use cases, so teams need not rebuild controls: identity context (who requested, under which role and scope); data boundaries (which sources a system may read and which targets it may write); policy evaluation checks run before retrieval, before tool calls and before output; lineage capturing versions of prompts, models, indexes and policies plus per-request trace IDs; and audit.
Mots-clés: Adnan Masood · AI Governance · Generative AI · Threat Modeling · Runtime Controls
萬安算交所 AIXX · AIXX · Bourse compute vérifiée · Accéder à la salle