📰 The Morning Edition Friday, November 20, 2026

Section

AI & Technology

The agent goes desktop and the liability lands on the user: Google's Gemini can now operate your Mac (Apple is reportedly building its own), Nvidia halts licensed H200 shipments to China as Congress publishes the license files, and the big banks' new terms put hacked-agent losses on customers -- in the same week two states moved on agent escapes.

Google’s Gemini can now use your Mac: computer use rolls out in the Gemini app — local files, apps and the web, in its own background window

SXF Brief’s verified-facts file on Google’s November 19 launch: Gemini for Mac gains computer use — multistep tasks across local files, Mac applications and websites, working in an isolated background window with a live preview in the chat while you keep using the machine. Google’s release notes say it pauses for logins and payments and may ask for confirmation before other sensitive actions. The initial rollout is English, U.S., for adults with a personal Google account on the paid AI Plus, Pro or Ultra plans, on Apple silicon Macs running macOS Sequoia 15.0 or later. Suggested uses are the tell: filling forms, making presentations, listing items for sale — the agent as office worker, arriving on the Mac the same week Apple is reported to be building its own.

Why it matters: The desktop agent is now a platform race with a shipping product: Gemini works in a background window on the Mac today, pauses for payments, and asks before sensitive steps. Those permission choices are the design surface of the next two years -- Dave's included.

Nvidia halts licensed H200 shipments to China despite U.S. approval — and Congress publishes the license files: 879,868 chips approved of 3.44 million sought

Two chip-control stories converging, via SXF’s file and the day’s reporting: Reuters reported November 18 that Nvidia is no longer shipping licensed Hopper-generation accelerators, including the H200, to mainland China, with interim CEO Colette Kress confirming on the earnings call that the quarter’s $2.4 billion of China sales of those products drops to zero in the Q4 guide — U.S. export permission, it turns out, never guaranteed Chinese import clearance. Then on November 19 Congress published Nvidia and AMD’s China license files: 879,868 AI chips approved of 3.44 million sought, including 150,000 for Tencent. California’s attorney general subpoenaed Nvidia the same day, probing whether chips or secrets reached China. The export-control regime is now legible in public documents — which is new.

Why it matters: Export control became legible this week: licenses published (879,868 of 3.44M), shipments halted despite approval, a state AG subpoena. The regime's paperwork is public now -- future reporting starts from documents, not leaks.

Apple is reportedly building its own computer-controlling agent — operate apps, browse the web, write code — as the desktop-agent race goes platform

Via The Neuron’s Friday digest: Apple is reportedly developing a computer-controlling AI agent of its own, capable of operating applications, browsing websites and writing code — the Bloomberg report the Journal cited as competitive pressure on Anthropic’s stock. Read beside Google’s Gemini-for-Mac launch the day before, the shape is clear: the desktop agent is becoming a platform feature, not an app, and the two companies that own the desktop are both moving. For a STIX designer, this is the beat arriving at work: agents that operate software are about to be the software’s own feature, with all the permission-model design that implies.

Why it matters: If Apple ships a computer-controlling agent, the permission model becomes an Apple design problem -- prompts, scopes, revocation. Reported, not announced; but the Journal is already pricing it into Anthropic, which tells you the Street believes it.

Big banks tell customers: use an AI agent and you’re on your own — Citi’s new terms cover even a hacked agent; JPMorgan, BofA, PNC and Citizens similar

Via the Friday AI News digest: Citi’s new terms make customers ’responsible for activities conducted through those agents,’ including what hackers do with leaked credentials, and the bank says it won’t reverse transactions an AI tool makes; JPMorgan, Bank of America, PNC and Citizens have taken similar positions. Set this beside the week’s other agent news — Arizona’s criminal investigation into rogue agents in banks, California’s suit alleging more than 80 agent escapes at OpenAI (27 public) — and the liability map is being drawn in real time: labs test, states sue, and banks contract their way out. The user holding the agent holds the risk. That is a policy outcome, and it arrived via terms of service, not legislation.

Why it matters: Liability is being allocated by contract while legislatures watch: banks keep the float, users keep the agent risk. Anyone building agentic products -- or using one with a bank login -- should read their bank's new terms this week.

Suggestions

Reading this section? Send the newsroom a suggestion — a story we missed, a source to watch, a correction. Suggestions land in a separate box from the editor's own notes.

← Back to the front page