AI news for Friday, October 23, 2026: Treasury proposes China AI investment ban, banks lock out agents

Washington and Wall Street drew new lines around AI on the same day, and agents land on the wrong side of both.

Policy Business3 min
Treasury proposes banning US investment in China AI
Treasury's Oct. 22 draft rule would bar US investment in the development of any AI system by companies tied to China and other countries of concern, agent wrappers included. The compute threshold is gone, and so is the safe harbor for small AI deals.
Why it matters If you invest in, partner with or fund an AI startup with ties to China, Russia or Iran, the old question was how big its model was. Under this draft it is whether it builds with AI at all, and an agent built on someone else's model counts. Read it before your next term sheet, and comment by Dec. 7 if it hits your deals.
SRCU.S. Department of the Treasury
Agents Safety3 min
Banks lock out AI agents as 1Password offers a key
Major US banks are blocking or limiting AI agents that sign in with customers' passwords, Axios reported, as 1Password unveiled a way for agents to log in without ever seeing one. Sharing your bank login with an agent is over. Delegated sign-in is next.
Why it matters If an agent like Muse has your bank password, take it back today: banks are rewriting their terms against it and you may eat any loss. If you build agents, stop storing user credentials. Per-login approval and read-only access are where banks and 1Password are both heading.
SRC1Password
Research Agents3 min
Microsoft's MAI-DxO beats doctors on hard cases in Science
Microsoft says its MAI-DxO diagnostic system got 83% of 304 complex NEJM cases right, against 36% for physicians with references, in a study published in Science. The orchestrator gets the headline. The model underneath does most of the work.
Why it matters If you build AI for high-stakes calls, copy the design: agents that debate, with a price on every extra step. But the gain mostly comes from the underlying model, the comparison left out specialists, and these are benchmark cases, not patients. Check claims like this against your own evals before you believe them.
SRCMicrosoft AI
Products Models3 min
OpenAI's Decisions API goes GA with 400 questions a call
OpenAI's Decisions API left beta on Oct. 22, now taking up to 400 typed questions per request, up from 200, at $0.10 per million uncached input tokens with free output and new caching for shared input. The badge is minor. The doubled question count is what cuts your bill.
Why it matters If you pay a chat model to route, tag or moderate, you can now ask 400 yes, no or pick-one questions of one document in a single call and pay only for input. It is not the top-ranked decision model, so test it against your own labels before you move production traffic.
SRCOpenAI
Business Agents3 min
Barclays to cut 5,000 jobs, developers included, citing AI
Barclays plans to cut about 5,000 roles, roughly 5% of its staff, by the end of 2026, with about 60% in support functions that include developers. It is pushing Claude Code to half its developers at the same time, and that pairing is the story.
Why it matters If you write code at a large company, your employer is likely tracking how much of your team uses coding agents. Barclays put a number on it and named developers in the same quarter it announced cuts. Get fluent with the agents and own the review and design work they cannot.
SRCEuronews

Prompt of the day

Audit every AI agent that can sign in as you

› I want to audit which AI tools and agents can act on my behalf, because banks are now blocking agents that log in with customers' passwords.

Here is every AI tool, agent, browser extension or automation I use that touches my accounts: [list each one, and what it connects to: email, bank, calendar, cloud drive, company apps].

For each one, tell me:
1. How it most likely gets access (my password typed or stored, OAuth or a "Sign in with" button, an API key, a browser session, or read-only data sharing like an account aggregator). Ask me if you can't tell.
2. What it can do with that access: read only, or act (send, buy, move money, delete).
3. A risk rating from 1 to 5, with one sentence on why, weighing what it can reach against what it actually needs.
4. The safer setup: a read-only permission, a scoped API key, approval before each sign-in, or removing access entirely.

Then give me:
- The three changes to make today, in order.
- Any account where I should change my password now because I shared it with a tool.
- A short checklist I can rerun every month.

Be direct. If a tool doesn't need the access it has, say so.

Open in Magai

Works in any AI chat. In Magai, run it against several models at once.

Get every edition

This one arrived in inboxes at 6am ET.

Treasury proposes banning US investment in China AI

Banks lock out AI agents as 1Password offers a key

Microsoft's MAI-DxO beats doctors on hard cases in Science

OpenAI's Decisions API goes GA with 400 questions a call

Barclays to cut 5,000 jobs, developers included, citing AI