- Roko's Basilisk
- Posts
- Who Sets The Spending Limit
Who Sets The Spending Limit
Plus: Anthropic names its attackers, California curbs kids' feeds, Altman floats a slowdown.
Here's what's on our plate today:
🧪 Meta's Muse shops through Stripe's Link, the wallet that sets its spending limit.
🗞️ Anthropic blocks Russian, Chinese attacks; California curbs kids' feeds; Altman weighs slowing AI.
💡 Roko's Pro Tip: give an agent a single-use card number, never your real one.
Let’s dive in. No floaties needed.

Label Faster. Train Smarter. Ship Better Models.
Multimodal models are only as good as the data behind them. Our trainers work across text, audio, and image pipelines—handling transcription, labeling, and annotation tasks with speed and consistency.
Audio transcription and speech data labeling
Image and video annotation for vision model pipelines
Multilingual coverage across Portuguese, Spanish, and English
Part of our vetted LATAM talent network, working in U.S.-aligned time zones.
*This is sponsored content

Goodies delivered straight into your inbox.
Get the chance to peek inside founders and leaders’ brains and see how they think about going from zero to 1 and beyond.
Join thousands of weekly readers at Google, OpenAI, Stripe, TikTok, Sequoia, and more.
Check all the tools and more here, and outperform the competition.
*This is sponsored content

The Laboratory
TL;DR
AI assistants can now shop for you, and nobody wants to hand them your card number.
The problem: Software that browses the open web cannot store the 16 digits on your card, since anyone who has them can charge them anywhere.
The fix: A wallet keeps the real card and issues the assistant a stand-in number good at one shop for one purchase.
The position: Whoever issues that number sits inside every purchase and decides what the assistant may spend.
Who took it: Meta's Muse launched on September 8, 2026, accepting payments through Stripe's Link, and Anthropic's shopping guide leaves payment out entirely.
The doubt: OpenAI owned the checkout in ChatGPT for six months, then handed it back to merchants.
The next AI battle is over who gets to control your money
In ancient Roman households, a slave could run a business. The head of the household, who owned the property and the people in it, would set aside a fund of coin, livestock, or goods and hand it over to be traded. The slave then went into the market to buy, sell, lend, and strike deals with merchants who had no dealings with the master at all. The fund was called the peculium, and although it remained the master's property in law, everyone involved treated it as the trader's working capital. The arrangement worked because it was bounded. A merchant who had been cheated by such a trader could sue the master, but only for as much as the fund was worth, which meant the master could send someone into the market without putting his house and land behind every bargain that person struck. He was letting another party spend his money without handing it over, and the difference between the two was the size of the fund.
Paying for something online has never needed that distinction. You type the 16 digits printed on your card into a shop's checkout, and the shop charges them. The arrangement holds because the person typing the number owns the account behind it, with the digits serving as the identifier and a code from the bank standing in for a trip to the branch. However, since the thing doing the typing is increasingly an AI assistant that shops on your behalf, those digits now sit with software that reads pages written by strangers and can be talked into things. Whoever holds them can charge them at any shop in the world until the card expires. That is the question this whole business now turns on, and it is the Roman one: how do you let software spend your money without giving software your money?
Everybody building these assistants has reached the same answer. Something trustworthy holds onto the real card and gives the assistant a stand-in number that works at one shop for one purchase, so the account stays where it is, and the assistant never sees it. That something is called a wallet, and it is worth following through the rest of this story. Answering the question this way puts whoever runs the wallet inside every purchase the assistant makes, and leaves them deciding how much it is allowed to spend. That could turn out to be a better place to stand than owning the assistant itself, and the week of September 8, 2026 showed who is standing there. Meta launched Muse that day, a personal assistant available in the US on iOS, Android, and WhatsApp, according to Meta's newsroom. Stripe, which handles card payments behind a large share of online shops, confirmed the same day that Muse pays through Link, the consumer wallet Stripe already runs. Six days earlier, Anthropic had published a guide for companies building shopping assistants on Claude, and it contains no payment system at all. Neither model company answered the question itself, and the answer is being supplied by firms that were in payments long before there were assistants to serve.
The stand-in number is the whole of the answer
The 16 digits on a card are called a primary account number, and replacing them with a limited stand-in is called tokenization. A stand-in is only as safe as the rules attached to it. Stripe's version can be used at a single-named shop, up to a set amount, for a short window, and it does not contain the real card number anywhere. An assistant can therefore pay for the thing in front of it and cannot go and buy something else afterward. Those rules are the Roman fund in modern form, drawn tight around a single order, and somebody has to write them on the shopper's behalf. That is where the question stops being technical, because the company issuing the stand-in is the company deciding what your assistant is allowed to buy.
Keeping the money out of the assistant's hands still leaves a second problem above it. The assistant, the shop, and the payment company have to agree on what is in the basket, what it costs, and what exactly the shopper said yes to, and a safer card number settles none of that. What they need is a shared format for passing that information back and forth, and two are competing for the job. The Agentic Commerce Protocol was written by Stripe and adopted by OpenAI and Meta, while the Universal Commerce Protocol was built by Shopify and Google and keeps the shop itself as the legal seller of the goods. These are two separate fights, and winning the one about how money moves does not win the one about how orders are described. That distinction is worth holding onto when working out what Meta actually got when it launched an assistant that can pay.
Stripe was inside the transaction before Meta arrived
Meta rented its answer rather than building one. Stripe's own announcement gives the order of events, because Link was already handling payments for two other assistants, Grok Bot and Instinct, which makes Muse the third name on the list rather than the first. Stripe has over 300M users and more than 1M businesses accepting it. Where a shop does not accept Link, Link issues Muse a one-time card number good only for the purchase the person approved. That approval matters as much as the card number does, since Meta says Muse cannot see your payment details and asks before it buys anything, and Stripe says the shopper confirms the total in the chat window. A person is still setting the limit, purchase by purchase.
Anthropic went a different way. Its guide covers how an assistant should read a shop's catalog, build a basket, remember what a customer likes, and look up past orders, and it leaves the payment itself to whichever company deploys the assistant or to an outside payments provider. No money passes through Anthropic at all. Its reported work on payments in-house is a separate thing, because The Information described job listings for billing, payments, fraud, and tax systems that charge customers for using Claude rather than anything touching a shopper's card, and Anthropic told the publication that Stripe remains a strong partner. Two of the largest model companies faced the same question in the same week, and both left the answering to somebody else.
The card networks and the bodies that write payment rules treated that gap as worth filling. Biometric Update reports that the FIDO Alliance, which sets standards for proving who you are online, is writing frameworks for this kind of shopping. It also reports that EMVCo, the body owned by the card companies, has published one covering payments made by assistants, with comments open until September 30. Both are moving to make sure that when an assistant pays, it still pays over the card rails they control.
The one company that owned the checkout gave it back
The strongest argument against treating this as the prize is that someone has already tried it and given up. OpenAI added a checkout feature to ChatGPT on September 29, 2025, calling it Instant Checkout, and released the format publicly with Stripe. Six months later, it changed course, telling CNBC in March 2026 that the first version was not flexible enough, so shops would go back to running their own checkouts while ChatGPT concentrated on helping people find products. Bob Hetu, an analyst at the research firm Gartner, told the network that OpenAI had underestimated how hard enabling transactions would be. Stripe was running the wallet underneath the whole time and it did not save the product, because what sank it was shoppers abandoning the purchase, sales tax being difficult to get right, and stock information that did not match what the shop actually had.
None of that means the wallet is worthless, because the version that failed was when Instant Checkout could only handle a single item at a time; it pulled product details from merchant websites rather than receiving them properly, and it had no system for calculating sales tax. Its collapse says more about that build than about a better one. Muse has not settled the question either, since an assistant that stops and asks before every purchase is really a faster checkout button, and the hard part only arrives when that confirmation disappears.
The Roman arrangement worked because the limit was set in advance and everybody knew where it sat. The emerging version asks the same thing of an assistant that can shop on its own: how much of your money should a machine be allowed to control without ever holding the money itself? The company that issues the stand-in number sits close to that decision, and today those companies are mostly the payment firms that built the rails before AI agents existed. Whether that becomes real power or remains plumbing will depend on what happens when nobody is confirming each purchase. Someone will have to decide in advance how much a machine is allowed to spend.


Roko's Pro Tip
![]() | 💡If you are wiring an agent into checkout, never hand it a live card. Issue a stand-in number scoped to one merchant, one amount, and a short window, and log who approved it. The limit you write in advance is the only control you keep once the confirmation step goes away. |

250,000 professionals read it before their first meeting.
Bay Area Times covers AI, startups, business, and the innovation economy in a format built for people with no time to spare.
One visual per story, concise analysis under it, and you are done in three minutes. Free, Monday to Friday, and read by founders, operators, and investors across the tech ecosystem.
If you only keep one daily newsletter, this is a strong candidate.
*This is sponsored content

Monday Poll
💳 Your AI assistant is about to buy something. Who should decide how much it can spend? |

Bite-sized Brains
Anthropic names its attackers: The company said it disrupted a suspected Russia-linked espionage campaign and seven China-based labs, including operators tied to Alibaba that ran more than 151M exchanges to distill Claude.
California caps addictive feeds for kids: Newsom signed 13 bills barring under-16s from infinite scroll and algorithmic autoplay, and banning toys with companion chatbots for four years.
Altman floats a coordinated slowdown: He told staff at a company meeting that OpenAI could pace frontier development alongside rival labs, while acknowledging some may not agree.
Meme Of The Day

The Toolkit
Dust: No-code platform for building custom AI agents that plug into your company's tools and automate workflows.
Krea: Real-time AI image and video generator built for designers who want to steer output, not fight prompts.
Continue: Open-source AI code assistant that plugs into VS Code and JetBrains with full control over models and context.

Rate This Edition
What did you think of today's email? |







