# How much permission should an AI assistant have? | AGI Huang argues that agentic systems should operate with restrictive rights. --- Canonical page: https://agiscorecard.com/future-guide/agent-permissions Machine-readable verdicts: https://agiscorecard.com/data.json (CC BY 4.0) This Markdown mirror is generated from the page; the HTML page is canonical. ## Source-linked future guide You can read all sources and guides without JavaScript. Enable it to filter, save or export your own plan. Future guide (/future-guide) / Give access deliberately Give access deliberately How much permission should an AI assistant have? Huang argues that agentic systems should operate with restrictive rights. Jensen Huang · CNBC Interview: 2026-09-28 Source review: 2026-10-03 CNBC · Video Jensen Huang Return to the original context. Watch at publisher ↗ (https://www.cnbc.com/video/2026/09/28/nvidia-ceo-jensen-huang-all-systems-around-ai-systems-must-be-designed-with-restrictive-rights.html) Publisher’s original page ↗ (https://pressroom.versantmedia.com/cnbc/press-releases/cnbc-excerpts-nvidia-founder-ceo-jensen-huang-speaks-cnbcs-squawk-box-today) Source, interpretation and uncertainty What does this actually establish? Source status Source locator: HUANG ON AGENTIC SYSTEM · Not an independently verified outcome Speaker context NVIDIA founder and CEO · AI infrastructure vendor perspective What remains open Restricting access reduces exposure but cannot establish that an entire system is safe. AGI editorial interpretation Two ways this could unfold Narrow permissions can contain some errors. A poorly chosen permission can still enable a costly action. These are conditional scenarios, not the speaker’s words or a probability estimate. Read alongside: Why can a capable AI still be hard to deploy? (/future-guide/deployable-trust) Different framing, dates and definitions; not a direct head-to-head forecast. Where do I stand for now? Lean toward agreement Lean toward disagreement Need more evidence This is your private, device-only record. No public vote count or collective probability is implied. Turn a view into a small experiment What is my next step? A simple plan you can test and revise. Your notes stay in this tab until you save them. What matters to me Understand the change Adapt my work Choose what to learn Explore income ideas Think about education Test my judgment The task or question in my life One action I will try Use the editorial starting point ↗ What would change my view? My review date (optional) I have tried this action Save my next step Export plan Add to calendar Saved on this device. You control the plan and review date; these are not model predictions. Read the interview and evidence on AGI ↗ (https://agiscorecard.com/future-guide/agent-permissions) Embed the free action planner Video and context stay on this source page. ← Return to my next steps (/future-guide#notebook) ✳ JARVIS Research guide What is the smallest permission set this task needs? Separate read access from consequential actions in one workflow and review each permission. Make my next step ↗ (#plan) Based on published sources and editorial guidance. This is not an AI-generated forecast or a personal capability assessment. A useful reading path A view in its original context A reason to stay open-minded An action you can try No account required. When you want to keep the work Your next step deserves a place to return to. Reading, video access, local saves and exports are free. Use the same plan on another device with AGI’s existing cloud workspace and version history. No paid interview collection or automated personal-report subscription is offered here. AGI cloud membership 9 USDT / 30 days Plus the order-matching decimal and network fees. 50 workspaces, 10 versions per workspace, 64 KB per workspace and 5 MB in total. No auto-renewal or extra AI allowance. Open my plan in membership The member page checks availability. Opening it does not upload your plan or start a purchase; save there explicitly. USDT payment only.